E2e Working Paper 007

How Consumers Respond to Environmental Certification and the Value of Energy Information

Sébastien Houde

Revised December 2016

This paper is part of the E2e Project Working Paper Series.

E2e is a joint initiative of the Energy Institute at Haas at the University of California, Berkeley, the Center for Energy and Environmental Policy Research (CEEPR) at the Massachusetts Institute of Technology, and the Energy Policy Institute at Chicago, University of Chicago. E2e is supported by a generous grant from The Alfred P. Sloan Foundation.

The views expressed in E2e working papers are those of the authors and do not necessarily reflect the views of the E2e Project. Working papers are circulated for discussion and comment purposes. They have not been peer reviewed.

Published in

The RAND Journal of Economics, 2018, 49(2), 453-477.

How Consumers Respond to Environmental Certificationand the Value of Energy Information

Sebastien Houde∗

August 1, 2014

Abstract

The ENERGY STAR certification is a voluntary labeling that favors the adoption of energyefficient products. In the US appliance market, the label is a coarse summary of otherwise readilyaccessible information. Using micro-data of the US refrigerator market, I develop a structuraldemand model and find that consumers respond to certification in different ways. Some consumershave a large willingness to pay for the label, well beyond the energy savings associated with certifiedproducts; others appear to pay attention to electricity costs, but not to the certification, and stillothers appear to be insensitive to both electricity costs and ENERGY STAR. The findings suggestthat the certification acts as a substitute for more accurate, but complex energy information. Usingthe structural model, I find that the opportunity cost of having imperfectly informed consumers inthe refrigerator market ranges from $12 to $17 per refrigerator sold.

JEL: Q41,Q50,L15,D12,D83.Keywords: Product Certification, Attention Allocation, Demand Estimation, ENERGY STAR.

∗An earlier version of this paper was circulated under the title: How Consumers Respond to Product Certifi-cation; A Welfare Analysis of the ENERGY STAR Program. I am indebted to Jim Sweeney, Wes Hartmann,Jon Levin, and John Weyant for all their helpful advice. I would also like to thank Larry Goulder, MattHarding, Ken Gillingham, Anant Sudarshan, Jim Sallee, Charles Mason, David Rapson, and seminar par-ticipants at the Stanford IO lunch, the SEEPAC environmental economics seminar, and the NBER SummerInstitute in Environmental and Energy Economics.Department of Agricultural and Resource Economics, University of Maryland. Email: shoude@umd.edu.

2

1. Introduction

The role of certification programs is to address informational market failures, and ultimately help

consumers in distinguishing high-quality products from lemons (Akerlof 1970; Viscusi 1978). The-

oretical analyses have shown that the welfare effects of certification programs depend on a con-

fluence of factors such as market structure, heterogeneity in consumer preferences, and consumer

sophistication (Dranove and Jin 2010). When certification programs come with labeling schemes,

the information content of labels may also be an important determinant of welfare. Labels—like

advertisements—may affect how consumers perceive and experience goods (Bagwell 2007). More-

over, whether a label provides new information, or summarizes complex, but otherwise available

information have different welfare implications.

The goal of this paper is to show how consumers respond to ENERGY STAR (ES), a voluntary

certification program managed by the US Environmental Protection Agency (EPA). The main

feature of the program consists of a simple labeling scheme (Figure 1(a)) that identifies the most

energy efficient products in the marketplace. In several markets, the program complements the

mandatory label EnergyGuide (Figure 1(b)), which provides detailed energy information. The

rationale of the ES label is to present energy efficiency as a dimension of quality that is simple,

easy to assess, and fully salient to consumers.

I propose a structural demand model that captures the various mechanisms by which the cer-

tification influences consumers. The model is inspired from classic models of consumer search and

information acquisition (e.g., Stigler 1961; McCall 1965; Gabaix et al. 2006). In the same spirit as

Sallee (2013), I use a theory of rational attention allocation (Sims 2003) to model why consumers

may rely on different pieces of energy information.

The estimation is carried on a large sample of transactions of the US refrigerator market. In

addition to rich variation in prices, electricity costs, and financial incentives for purchasing certified

products, the sample period includes two natural experiments that provide variation in labeling.

The structural model exploits this variation to recover consumer preferences for the ES label,

marginal utility of income, discount rate, and propensity to take advantage of rebates. The crucial

feature of the model is, however, to recover heterogeneity in the way consumers rely on energy

3

information. I find that some consumers rely on the certification and value the label itself well

beyond the energy savings associated with certified products, but others rely on a measure of

electricity costs and do not value the label highly. A non-negligible fraction of consumers also

appears to neither value the certification nor consider electricity costs in their purchase decisions,

and this fraction is much larger for lower income households.

The results from the structural estimation are supported by alternative estimators that account

for heterogeneity in a more flexible manner. I first show using a simple conditional logit that the

coefficients on the ES label and electricity costs are inversely correlated with several demographic

variables. In particular, richer and smaller households, where the head is older and educated

respond more to electricity costs, but not so much to the label. Using an approach proposed by

Fox et al. (2011), I recover a nonparametric join distribution of the coefficients on the ES label

and electricity costs, which also shows a clear negative correlation between the two coefficients.

Altogether, the data suggest that the certification acts as a substitute for more accurate, but

complex energy information.

These results are of primary importance for the design of energy policies that target energy

demand. Consumers’ inattention to energy efficiency is an often-cited justification for minimum

standards. Recent empirical evidence has, however, been mostly confined to the car market (e.g.,

Klier and Linn 2010; Li, Timmins and von Haefen 2009; Busse, Knittel and Zettelmeyer 2013;

Allcott and Wozny 2012), and suggests that inattention to car operating costs may in fact be

modest on average. I fill a gap in this literature by explicitly showing how existing labeling policies

in the appliance market may overcome inattention, and quantifying the value of energy information.

I find that the opportunity cost of having imperfectly informed consumers in the refrigerator market

ranges from $12 to $17 per refrigerator sold, or $108M/year to $153M/year for the whole market.

These latter estimates are more than twice larger the overall program cost of the ES program.

Surprisingly, although ecolabels have become ubiquitous in marketplaces, empirical analyses

of such policies have remained scarce (Mason 2013). The structural model contributes to the

theoretical work on the topic by developing and testing a theory of how consumers respond to

4

environmental certification. It also provides the first estimates of the behavioral effects of the ES

program using a revealed preference approach combined with market-level data.1

More broadly, this paper contributes to the literature on information-based policies. Arguably,

the complexity and economic significance of the appliance purchasing decision pale by comparison

to other decisions such as retirement savings, health care plan, nutrition, and school enrollment.

In these settings, it is also customary to provide coarse summaries of complex information, by way

of grades, rankings, and score cards, for instance. My framework and empirical strategy could be

applied to these contexts, and complement other structural estimators (e.g., Abaluck and Gruber

2011; Abaluck 2011; Handel and Kolstad 2013) that aim to explicitly deal with attention allocation

and heuristics used in complex choice environments.

The remainder of this paper is organized as follows. Section 2 provides a primer on energy

labels for the US appliance market. Section 3 presents a theory of how consumers respond to the

ES certification. Section 4 gives an overview of the data, and discusses sources of identification.

Section 5 provides preliminary estimates using a simple conditional logit model. Section 6 proceeds

to the estimation of the structural model, and section 7 provides an alternative estimator that

recovers heterogeneity in the valuation of energy efficiency with minimal structure. The policy

analysis is presented in section 8, and conclusions follow.

2. A Primer on Energy Labels for the Appliance Market

In the US appliance market, two different labeling schemes offer energy information to consumers.

The EnergyGuide program was established in 1979, and is managed by the Federal Trade Com-

mission. The program covers most types of appliances, and several consumer electronics. The

main feature of the program is the requirement that all products offered in the marketplace display

prominently a large yellow label with detailed technical information about energy consumption and

costs. The content of the EnergyGuide label was revised several times since the inception of the

program. In its most recent version (Figure 1(b)), the label displays the annual electricity (or gas)1Ward et al. (2011) used an online survey to elicit consumers’ willingness to pay for the ES label, and found apositive and significant impact of the label on hypothetical decisions. Recently, Newell and Siikamaki (2013)use an online survey with hypothetical choices and also found that the label has a large impact on choices.

5

consumption of the product, an estimate of the annual operating cost based on the national average

price of energy, and a scale with a cost range of similar products.

ENERGY STAR was established in 1992 by the EPA as a way to complement EnergyGuide.

The ES label contains no technical information, and is similar to a simple brand logo (Figure 1(a)).

Only products that meet a certain certification have the right to display the label. For certified

appliances, the label is often displayed on the product itself, and on the EnergyGuide label.

Historically, the certification requirement for appliances has been binary.2 Under this system,

the requirement is defined relative to the federal minimum energy efficiency standard. For instance,

since April 2008, a full-size refrigerator can be certified if it consumes at least 20% less electricity

than the minimum energy efficiency standard established for this refrigerator model. Because

minimum standards vary as a function of the appliance size, and other attributes, the certification

requirement does not fix electricity consumption to an absolute level. It is common to find small

uncertified appliances that consume less than larger certified appliances.

A crucial feature of the program is that the certification requirements are revised periodically.

The EPA revises the stringency of a requirement using various criteria, such as the proportion of

certified products offered on the market, the market shares, and the availability of new technologies

(McWhinney et al. 2005). The stringency of a requirement is ultimately determined by the EPA

upon consultation with different stakeholders, such as manufacturers, part providers, retailers,

analysts, and environmental groups. The EPA usually announces a revision one year in advance.

An important institutional feature of the US electricity market is that several electric utilities are

subject to regulations that incentivize them to promote energy efficiency measures to consumers.

Rebate programs tied to the purchase of ES products have been arguably one of the most pop-

ular energy efficiency measures pursued by US electric utilities. The Energy Policy Act of 2005

also established a state rebate program for appliances tied to the ES certification. In 2010-2011,

this program was temporarily funded with stimulus funds. Thus, apart from the informational

component, financial incentives may play a role in favoring the adoption of certified products.2Since 2011, the EPA has been experimenting with a multiple tiers system in some product categories. Thepresent paper focuses on the binary system.

6

3. An Information Acquisition Model for Energy-Intensive Durables

In the context of purchasing an appliance, evaluating the energy cost of each product is arguably a

complex and costly task. It requires taking the time to look at the EnergyGuide label, understand

the various pieces of information, look up the local electricity prices, and perform mental calcula-

tions. A priori, knowledge about the energy cost of each product j, Cj , is then consider imperfect.

A consumer i can learn the value of Cj only by incurring a cost Ki(ei), where ei represents the

effort to collect and process energy information. Define I(ei) as the consumer’s knowledge about

energy costs for a given level of effort, the expected energy cost of product j at the time of purchase

is then defined by: E[Cj |I(ei)].

In addition of the expected energy cost, consumer i values product j as a function of its qual-

ity (δj), price (Pj), whether a product is ES certified or not (Dj), rebate amount (R), and an

idiosyncratic taste parameter (εij):

(1) Uij = δj − ηPj − θE[Cj |I(ei)] + τDj + ψDj ×R+ εij ,

where the coefficients: η, θ, τ , and ψ all relate to structural demand parameters. The coefficient

on prices (η) corresponds to the marginal utility of income, and the ratio of the coefficient on the

ES dummy (τ) and the marginal utility of income (τ/η) corresponds to the marginal willingness

to pay for the label itself. If we were able to observe whether a consumer claimed a rebate R, and

there were no hassle costs associated with claiming a rebate, the coefficient on rebates would match

the coefficient on prices, i.e., would be equal to the marginal utility of income. These conditions

do not hold in the present application. The coefficient on rebates (ψ) should then be interpreted

as an intent to treat estimator (Imbens and Angrist 1994) corresponding to the marginal utility of

income multiplied by the probability that a consumer takes advantage of a rebate program. Finally,

comparing the coefficient on electricity costs (θ) with the coefficient on prices informs about the

extent that consumers trade off future electricity costs with the purchase price. Assuming that

consumers form time-unvarying expectations about annual electricity costs, and do not account for

the effect of depreciation, the coefficient on electricity costs is then a reduced form parameter that

7

relates to the discount factor as follows:3

(2) θ = η1− ρL

1− ρ ,

where L is the lifetime of the durable, and ρ = 1/(1 + r) is the discount factor. The estimates of η

and θ can then be used to infer a value of an implicit discount rate r.

A consumer’s purchasing decision is modeled as a two-step process. To begin, a consumer is

uncertain about energy costs, the meaning of the certification (noted S), and whether a product is

certified or not.4 A consumers observes the quality and the price for each appliance and then decides

the amount of effort to collect and process energy information. I assume that choosing the effort

level is a discrete choice with three possible outcomes. A consumer may collect and processes enough

information to form accurate expectations about the energy costs associated with each option, and

thus make a purchase decision perfectly informed (noted e = I). In this case, E[Cj |I(ei = I)] = Cj ,

where Cj is an accurate forecast of the energy costs of product j. Alternatively, a consumer may

only collect information related to the certification (noted e = ES). He will then learn the true

meaning of the certification and which products are certified, but not the exact energy costs of

each option. The expected energy costs are then: E[Cj |I(ei = ES)] = E[Cj(s)|Dj ], where s

corresponds to the certification requirement.5 Finally, a consumer may decide to not update his3Under these assumptions, the lifetime energy costs (LCj) over the lifetime of the durable are given by:

LCj =L∑t=1

ρtCj = 1− ρL

1− ρ Cj .

4This can be interpreted as a small psychological cost to detect whether or not a product has the ES label.5If a consumer were to know the meaning of the ES certification, he would then know that the energy costs ofcertified products are below a particular threshold corresponding to the ES standard. Denote this thresholdby S. If the consumer does not collect information, his belief about the meaning of S is given by a priordistribution G. If the consumer learns the meaning of the certification, he observes a realization s from G.If the prior on energy costs is given by the distribution F , the expected energy cost for a certified productj is then:

E[Cj(s)|Dj = 1] = E[Cj |Cj ≤ s] = 1F (s)

∫ s

c

cf(c)dc,

where c is the lower bound of the support of F , and . Similarly, the expected energy costs for a non-certifiedproduct j′ is: Similarly, the expected energy costs for non-certified models are:

E[Cj′(s)|Dj = 0] = E[Cj′ |Cj′ > s] = 11− F (s)

∫ c

s

cf(c)dc,

where c is the higher bound of the support of F .

8

priors and remain uncertain about energy costs, the meaning of the certification, and whether a

product is certified or not (noted e = U). In this case, prior beliefs are uninformative, i.e., all

options in the choice set have the same expected energy cost ex ante: E[Cj |I(ei = U)] = C. In a

discrete choice framework, it is as if the consumer dismisses energy costs altogether.

Prior to making a purchase, the consumer’s effort to collect and process information is given by

the following optimization problem:

(3) maxe={U,ES,I}

−Ki(e) + Eε,D,S,C

[maxj{Uij(δj , Pj , Cj(S), Dj , εij)}|I(e)

].

When consumers are heterogeneous with respect to the costs of collecting and processing energy

information, a fraction of the population will rely on energy costs, another on the ES certification,

and others will dismiss the energy efficiency attribute.

In this framework, the certification influences consumers via three mechanisms. First, it can be

used as a heuristic to compare products in a binary manner along the energy dimension, and thus

partly informs about energy costs. In particular, if consumers rely on ES, instead of an accurate

forecast of energy cost, they would value certified products more based on their expected energy

savings: E[Cj(s)|Dj = 1] − E[Cj′(s)|Dj = 0]. According to this mechanism, the certification

thus brings welfare gains because it allows consumers to economize on information costs while still

bringing some information about energy efficiency.6

A second mechanism is the effect of the label itself. That is, consumers might value certified

products beyond purely their expected energy savings, and value the exact same appliance with and

without the label differently. There are a number of reasons that can explain why the label might

have such effect. ES products are often advertised as being environmentally friendly. Consumers

might thus derive warm glow to purchase certified products, respond to social norms, or experience

social prestige. Under this interpretation, the label impacts how consumers experience the product,

and provides a positive utility shock. It is also possible that consumers correlate certified products

with better quality, and are subject to what marketeers refer to the halo effect (Boatwright, Kalra,6Consumers at the source of the welfare gains are the ones that would not consider energy efficiency ifthe certification were not in effect, but rely on the ENERGY STAR label when it is present. Note that aconsumer should always choose to be fully informed if there are no extra costs to doing so, i.e., a consumeris better off with more information. A formal proof of this argument can be found in Appendix A.

9

and Zhang 2008). In this case, the certification provides an illusion of higher quality, and biases

consumers. Empirically, these explanations are hard to disentangle, but they are important to

recognize as they have different welfare implications.

The third mechanism is simply the effect of rebates, which provide an explicit financial motiva-

tion to adopt ES products.

4. Data and Preliminary Evidence

The primary data used for the estimation consist of all transactions made at a large US appliance

retailer during the period 2007-2011 where a full-size refrigerator was bought. The retailer offers a

large selection of appliance models, and has at least one brick-and-mortar store in each US state

in addition to a national online store. For each transaction, I observe the date, the model of the

refrigerator, attributes, the suggested retail price, the promotional price, the wholesale price, the

taxes paid, and the zipcode of the store where the transaction was made. For a large subset of

transactions, I also observe consumer demographics, such as household size, income, education,

homeownership, housing type, political orientation, and age of the head of the household. This

demographic information is transaction specific.

The estimation focuses on refrigerators for three reasons. First, during this period, there were

two events that led to the decertification of ES refrigerators. As it is further explained below, these

events provide natural experiments that identify the effect of the label. Second, refrigerators are

one of the few energy intensive durables for which utilization behavior has little impact on energy

consumption. Operating costs can then be estimated accurately using engineering information

and electricity prices. When utilization is unobserved, this creates a measurement error problem

(Allcott and Wozny 2012), and can be at the source of a selection bias (Bento, Li and Roth 2012). In

the present application, I can credibly rule out utilization as a source of unobserved heterogeneity,

which has important implications on how to interpret the results. Finally, a third reason is that this

market is representative of other appliance markets in terms of market structure and public policies

it is subject to. The structural demand parameters can then be applied to a broader context.

10

The construction of the key variables used in the estimation, and their main sources of variation

are discussed next.

4.1. ENERGY STAR Certification

When a new certification requirement comes into effect, the EPA requires that manufacturers and

retailers remove the ES label on certified products that do not meet the more stringent requirement.

For products that display the ES logo on the EnergyGuide label, firms are also required to clearly

identify that these products are decertified by masking the logo. Using data that cover the period

before and after the revision in the standard, it is then possible to observe the same refrigerator

model being sold at the same store, with and without the ES label. I can thus exploit this variation

in labeling to identify how consumers value the ES label, while controlling for all other product

attributes.

The last revision of the ES certification requirement for full-size refrigerators came into effect

on April 28, 2008, and was announced exactly one year in advance by the EPA. Prior this revision,

ES refrigerators had to consume at least 15% less electricity than the minimum energy efficiency

standard; the revision set the stringency at 20%.

For the year 2008, there were 2,524 refrigerator models available in the whole US market and

1,278 models lost their certification (Table 1). In the sample, there are 1,483 models for that year,

with 674 models that were decertified. The large number of decertified models provides good power

for identification, and more importantly covers all product classes. This decertification event thus

allows us to elicit preferences for all segments of the population, but it brings two challenges.

First, manufacturers and retailers adjust to the revision in certification by introducing new

models, and removing decertified models. This is illustrated on Figure 2, which plots the location

of all the models offered for particular types of full-size refrigerators, with respect to the minimum

and ES standards. The two standards are depicted by the straight lines as a function of the

refrigerator size. Each dot represents a refrigerator model offered in a given year. There are two

important stylized facts. First, there is strong bunching at the standards. Firms thus believe

that consumers value the certification, and make product line decisions accordingly. Second, firms

can adjust their product lines quickly. In the same year that the revision was announced, models

11

just meeting the new certification entered the market. Following the revision, a large number of

decertified models also exited the market.

If left unaccounted for, product exit would confound the effect of the change in labelling. To

control for the effects of product entry and exit, I construct an unbalanced panel of refrigerator

models offered on the market. In particular, for each refrigerator model, I identify the first and

last month for which a model was sold at least once in a specific store. I assume that consumers

shopping at this store between these two months could purchase this particular refrigerator model.7

Under these assumptions, the choice set imputed for each consumer consists of refrigerator models

that were physically present in the store at the time of purchase.

The second challenge in using this large decertification event is that I do not observe the exact

date that the ES label was removed or masked on each product. Although, the retailer’s policy

is to coordinate the change in labeling across all stores, and to implement the change close to the

date mandated by the EPA, I cannot ensure that all store managers complied with the policy at

an exact date. The estimation is performed under the assumption that all refrigerators that were

decertified in 2008 lost their ES label on the last week of April 2008.

As a robustness test, I exploit a second decertification event that overcomes these issues. In

January 2010, the EPA announced that 21 refrigerator models were subject to an incorrect testing

procedure, which resulted in the underestimation of their electricity consumption. The revised test

confirmed that these models did not meet the certification requirement, and should be decertified.

The sample contains 14 of these models. Unlike the 2008 revision, this event came as a surprise

for manufacturers and retailers. The EPA also mandated to print new EnergyGuide labels printed,

and to diligently remove all references to ES after the public announcement on January 20, 2010.

This event is thus appealing because it provides a sharp discontinuity in labelling, and rules out

the effect of anticipation by the firms. One caveat is that the prices of these decertified models are

higher that the average price (Table 1). Therefore, if different consumers sort into product classes7Excluding the first and last month of the panel is a conservative assumption. This ensures that monthswere a stockout happened at the end or beginning of the month are discarded. Note that using store specificchoice sets is also a conservative assumption given that the retailer can deliver appliances available at nearbystores.

12

with expensive models, the behavioral response to this event might not be representative of the

whole market.

4.2. Prices

The retailer implements promotions primarily at the national level due to a national price policy.

Each refrigerator model is subject to large and frequent variations in prices. This high frequency

temporal variation in retail prices is exploited to identify consumers’ sensitivity to prices. Prices are

imputed from the transaction level data. Weekly and store-specific price time series are constructed

for each model, and the average tax rate is computed for each store and trimester.

Figure 3 plots the percentage change of average weekly prices relative to the suggested manu-

facturer retail price (a measure of promotion) together with the percentage change in weekly sales

relative to average weekly sales for each refrigerator model in the sample. Figure 3 suggests that

most of the promotions consist of price reductions between 5% and 25%, and that correlation be-

tween sales and prices is negative and relatively large. The figure also suggests a change in the

pricing strategy over time. In 2008, promotions tend to be more idiosyncratic, while in 2010 pro-

motions are clustered at rounded percentages (5%, 10%, 15%, and 20%). The average duration of

a promotion also increases from 2008 to 2010, from 1.4 week to 3.1 week (Table 1).

Figure 4 shows the weekly average normalized price for different classes of refrigerators.8 Around

the time of the revision in April 2008, there is no significant change in prices for decertified models

relative to models that were never certified, or met the more stringent requirement (i.e., models

that were certified before and after April 2008). Post-decertification, prices of decertified models

dip below other models by about 1% to 2%, but otherwise follows a similar trend. For 2010, the

story is different. The price of decertified models drop sharply just after the announcement by as

much of 15%.8The weekly average normalized price for different classes of refrigerators is computed as follows. First,the normalized price for each refrigerator model is obtained by taking the weekly retail price divided by itsaverage weekly price. The average normalized price and standard errors in each efficiency class are thencomputed by fitting a cubic spline on normalized prices.

13

4.3. Electricity Prices

The electricity cost of each refrigerator model is the yearly electricity consumption reported by

the manufacturer multiplied by the average electricity price of the region where each household

lives.9 I assume that consumers form time-unvarying expectations about electricity prices using

the current local average price. The time-unvarying assumption is consistent with recent evidence

in the car market suggesting that consumers’ best forecasts of future gasoline prices are simply the

current prices (Anderson et al. 2011). Note also that electricity prices tend to vary much less over

time than gasoline prices. For instance, between 2008 and 2010, the national average electricity

price remained virtually unchanged (Table 1). Whether consumers respond to marginal or average

electricity prices, and the appropriate level of spatial aggregation to compute average electricity

prices is open to discussions. Ito (2012) shows that consumers respond to variations in average

electricity prices within California, which suggests that fairly local average electricity prices is the

most appropriate measure. Most online tools provides by the US EPA and retailers to estimate

energy costs and savings associated with ES, however, rely on a state average. In the estimation,

I compute average electricity prices using two different levels of aggregation, namely at the state

and county level.

Average electricity prices are computed using the EIA-861 form of the Energy Information

Administration (EIA 2008), which provides total revenue and sales to residential consumers for

each utility operating in the US. Further assumptions and methodology to construct the average

electricity prices are described in Appendix B.

The main source of identification of the coefficient on electricity costs comes from the fact

that I observe the same refrigerator models being sold at stores located in different electric utility

territories. This allows me to control for product and region fixed effects and use cross-sectional

variation in average electricity prices across regions to identify the sensitivity to electricity prices.9I do not observe the zipcode of each household, but the zipcode of the store where each transaction wasmade. Each household’s location is imputed from the zipcode where the purchase was made.

14

4.4. Rebates

In the US, several electric utilities offer rebate programs to encourage the adoption of energy

efficient appliances. These rebate programs are all similar in nature. Consumers claim the rebates

by filling out a form that must be submitted by mail or online. The purchased refrigerator must

meet a given energy efficiency criterion, which for most programs consists of the ES certification. A

complete description of these programs is available in the database of state incentives for renewable

and efficiency (IREC 2011). The number of active rebate programs and the amount offered by each

program vary over time. In 2008, 87 utilities offered a rebate program for ES refrigerators, and this

number increased to 133 in 2010 (Table 1).

In addition to utility rebate programs, state governments also offered financial incentives for ES

appliances during the sample period. In 2010, as part of the stimulus effort, the federal government

allocated funds to all US states and territories to establish temporary rebate programs to encourage

the adoption of energy efficient products. Each state had sovereignty over the design of its rebate

program, which led to important variation in the rebate amount across regions and time. During

the 2010-2011 period, 44 states offered rebates for refrigerators. The average rebate amount was

$128, and varied from $50 to $600. Some programs lasted for only one day (e.g., IL and TX); the

longest running program was in Alaska and lasted 639 days (Aldy and Houde 2014).

In the estimation, the rebate amount offered to each consumer is the average rebate amount in

each county, which includes the utility and state rebates, when both are available. Utility rebates

vary across county and years, while state rebates have weekly variation.

5. The Average Consumer

The goal of this section is to characterize the preferences of the average consumer in the US

refrigerator market, and discuss the sources of variation used for identification.

The empirical strategy relies on a simple discrete choice model. Consumer i living in region r

at time t derives utility Uijrt from purchasing refrigerator model j:

(4) Uijrt = τDjt − ηPjrt + ψRrt ×Djt − θCjrt + δj + εijrt,

15

where Djt is a dummy variable that takes the value one if product j is certified at time t and zero

otherwise. The variable Cjrt is the annual electricity cost of product j sold in region r. Pjrt is the

retail price, which includes all the local sales taxes, and Rrt is the rebate amount offered in region

r. The variables γj , and εijrt denote respectively the product fixed effects, and the idiosyncratic

taste parameters.

The idiosyncratic taste parameters are i.i.d. and Type I extreme value distributed, which yields

closed-form expressions for the choice probabilities. The choice probabilities do not include an

outside option (no purchase or purchase at other stores) because the policy analysis focuses on

how energy information influences which product to purchase, not the timing of the decision. The

model is estimated by maximum likelihood, where the likelihood is constructed with the choice

probabilities of a large sub-sample of consumers. The sub-sample is taken from the transactions

that contain demographic information. I also only consider transactions made by homeowners

living in single family housing units that bought no more than one refrigerator in any given year.

This restricted set of transactions is then likely to correspond to households that have to pay for

their electricity bills. The sample thus rules out heterogeneity in the sensitivity to energy costs

due to the split incentive problem (Blumstein et al. 1980), i.e., the fact that some consumers of

energy-intensive durables do not pay for energy costs.

For the estimation, I treat the two decertification events distinctly. Moreover, because the

rebate programs offered as part of the stimulus effort are in nature different from the programs

offered by utilities, rebate amounts offered in 2010-2011 may have a different effect than rebates

offered in prior years. The estimation is therefore carried on two separate sub-samples. The first

sub-sample is taken from transactions close to the 2008 revision. Namely, only transactions from

January 2008 to December 2008 are considered.10 An additional rationale to consider a restricted

sample centered around April 2008 is to control for product exit. The second sub-sample is taken

from all transactions made between January 2009 and December 2011. This covers the second

decertification event and a period with more variation in rebates.10Transactions for the whole month of September 2008 and prior to January 2008 do not include demographicinformation. They are then excluded from the sample for the structural estimation. Data for the year 2007are used for robustness tests presented in the Online Appendix C.

16

5.1. Results

Table 2 presents the estimates obtained with various specifications of the conditional logit. For all

models, robust standard errors clustered at the store level are reported. The first three models are

estimated using the sub-sample covering the year 2008. Model 1 is the simplest and only controls

for product fixed effects. The coefficient on price is negative, relatively large, and significant.

The own-price elasticity evaluated at the mean price ($1300) is -5.47. This is large, but not

surprising. Each refrigerator model has several close substitutes, and manufacturers and retailers

offer generous promotions. Consumers can then easily substitute across models to take advantage

of the promotions.

The effect of the ES label is positive and statistically significant. This translates into a willingness

to pay for the label of $19, or about 1.5% of the purchase price, on average.

The coefficient on electricity costs is negative and significant. This suggests that consumers do

respond to information about electricity costs, on average. Assuming a lifetime of 18 years, the

implicit discount rate is 62%. This is on par with previous estimates for refrigerators that range

from 35% to more than 60% (Meier and Whittier 1983). To put these numbers in perspective, in its

latest cost-benefit analysis of minimum energy efficiency standards for refrigerators, the Department

of Energy (DOE) used a product lifetime of 19 years, and discount rates of 3% and 7% (DOE 2011).

The impact of ES rebates is positive, small, and not statistically significant. As explained earlier,

the coefficient on rebates should be interpreted as the marginal utility of income multiplied by the

probability that a consumer takes advantage of the rebate. For the present estimates of η and ψ,

the implied probability of taking a rebate is below 1.0%. The fact that the sub-sample covers the

year 2008 implies that only cross-sectional variation in utility rebates is exploited. This limited

source of variation combined with measurement error associated with using average rebate at the

county level may lead to an attenuation bias.

5.2. Other Specifications

Weekly variation in prices is exploited to identify the price coefficient. It can be argued that this

variation is in part exogenous to demand shocks due to some institutional details of the appliance

17

market. In this market, manufacturers set high MSRPs, and let retailers offer promotions. To

comply with antitrust laws, retailers, however, do not offer products in constant discount, and

cycle promotions across similar products. This pricing strategy provides idiosyncratic variation to

identify the coefficient on prices. At specific time of the year, the retailer, however, often offer

generous promotions for specific brands. Model 2 thus includes brand-week fixed effects to account

for marketing efforts correlated with promotions. This has few impact on the coefficient on prices,

and other coefficients as well, suggesting that the marketing effects are not a major source of bias.

The construction of the electricity operating costs relies on the assumptions that consumers do

not forecast future electricity prices, and use local average prices. The lack of variation in electricity

prices over time makes the first assumption uncontroversial. The second assumption may imply a

degree of sophistication that not all consumers might have. Measurement error in electricity prices

is therefore a concern. In the car market, Allcott and Wozny (2012) shows that measurement

error in car operating costs is an important source of bias. An important difference in the present

context is that the source of measurement error is not due to the unobserved utilization decision,

but due to not knowing which electricity price consumers rely on when making decision. Model 3

reports results where the average state electricity price is used, instead of a county average. The

magnitude of the coefficient on electricity costs increases (in absolute value), which suggests that

measurement error does have an effect. In this specification, the implicit discount rate is 32%,

about twice smaller than in the previous specifications. Using county average electricity prices thus

provides a conservative estimate of the average valuation of energy efficiency.

Because variation in electricity prices is primarily cross-sectional, a concern is that consumers

living in regions subject to higher electricity prices are systematically different than others. The

Online Appendix C contains additional specifications to control for region specific unobservables.

Count models that include brand-week fixed effects, product fixed effects, and store fixed effects are

presented. For all specifications, the coefficient on electricity prices remain negative and significant.

Model 4 is a conditional logit with product fixed effects estimated on data covering the period

2009-2011. This provides an alternative estimate of the effect of the ES label exploiting the 2010

decertification event. As discussed before, this decertification was unexpected by firms, ruling out

the possible confounding effect of product entry and exit associated with the 2008 revision. The

18

coefficient for the ES label is now more than twice larger than the coefficient obtained for the 2008

revision. The larger estimate can be explained in part by the fact the decertified models were

expensive, and thus captures preferences of affluent households that might have a high willingness

to pay for certified products. This previews the results of the next section, which show that the

willingness to pay for ES increases with income. Model 5 is similar to Model 4, except that it

uses only four weeks of data following the announcement on January 20, 2010. The goal of this

specification is to show that even with a more conservative criterion to rule out the effect of firms’

product line and inventory decisions, the coefficient on ES is still large. However, the estimate is

smaller than in Model 4, and imprecisely estimated. This reassures that the change in labeling does

capture demand preferences, and the label effect is not entirely confounded by firms’ decisions.

Finally, note that for Models 4 and 5, the coefficient on rebate is positive, but still economically

small, and only marginally significant (Model 5). The implied probability of taking advantage of

a rebate varies between 9% and 14%. These results are consistent with the findings of Aldy and

Houde (2014), which suggest that most consumers that took advantage of rebates in 2010 would

have purchased an ES appliance even in the absence of a rebate program.

5.3. Interaction with Demographics

I now let the demand preferences vary as a function of demographics. Table 3 first presents

the estimates from a conditional logit estimated for three different income groups. I distinguish

between households with income of less than $50,000, households with income between $50,000

and $100,000, and households with income of more than $100,000. As expected, lower income

households have a larger coefficient on prices, which implies that the marginal utility of income is

decreasing with income. The willingness to pay for the ES label is increasing with income, but the

implicit discount rate is decreasing. Everything else being equal, lower income households respond

less to the certification and electricity costs.

Table 3 also presents the estimates from a conditional logit estimated for the three different

income groups where the ES dummy is interacted with demographics. Doing so, the coefficient

on electricity costs remains virtually unchanged. This suggests that region-specific preferences for

energy efficiency are not an important source of bias for the coefficient on electricity costs. It would

19

be particularly problematic if, for instance, consumers living in high electricity prices regions also

tended to be greener and, therefore, valued ES more. In that case, high rates of adoption of ES

products in these regions would be confounded with a response to electricity prices. Interacting

the ES dummy with demographic information (income, education, political orientation, age of the

head of the household, and family size) provides a simple way to show that a positive correlation in

the preferences for the certification and electricity costs does not mechanically drive the behavioral

response to electricity costs.

In Table 4, both the coefficients on the ES dummy and electricity costs are interacted with

demographics. First note that within each income group, income is still an important source

of heterogeneity11 for the ES coefficient, except for the highest income group. Within the low

and medium income groups, the coefficient on ES increases with income. This suggests that the

willingness to pay for ES increases steadily with income until it reaches a certain threshold. The

coefficient on electricity costs increases with income within the low and medium income groups.

Here, a larger and positive coefficient implies that consumers value future electricity costs less (i.e.,

higher implicit discount rate). For these two income groups, there is then an inverse correlation

between the willingness to pay for the ES label and lower electricity costs.

The effect of education varies between income groups. For the two highest income groups,

households where the head has a graduate degree tend to respond more to electricity costs relative

to less educated households. On the other hand, these households do not value the ES label much

more. For the lowest income group, this is the inverse pattern. It thus appears that more affluent

and educated households might be sophisticated enough to consider electricity costs, and by doing

so they do not see extra value in the ES label.

The most important source of heterogeneity for the coefficient on electricity costs is, however,

the age of the head of the household. Older households respond more to electricity costs, and this

holds across the three income groups. The coefficient on ES has no clear correlation with age, and

the effects are marginally significant.11Income is measured with a categorical variable taking nine values. Each sub-sample corresponding to aparticular income group contains three income groups.

20

The size of the household matters, especially for the first two income groups. Larger families

value ES more, but respond less to electricity costs.

Finally, political orientation impacts mostly preferences for ES. Perhaps surprisingly, relative to

Republicans, households registered as Democrats or other political parties (including non-registered

households) have a lower willingness to pay for the label.

The Online Appendix C presents an additional model where the coefficient on prices is also

interacted with demographic information. Under this specification, the patterns discussed above

still hold.

5.4. Discussion

The aforementioned estimates suggest that both the ES certification and information about elec-

tricity costs impact consumers, on average. The sign and the magnitude of the coefficients for

the ES label and electricity costs are robust to several specifications. Measurement error in aver-

age electricity prices has however a relatively large impact on the coefficient on electricity costs.

Overall, consumers, at least some, appear to rely on different pieces of energy information when

purchasing an appliance.

Interactions with demographics information yield interesting heterogeneity patterns. The most

important finding is that the coefficients on ES and electricity costs are inversely correlated for most

demographic variables. This implies that classes of consumers that value ES highly tend to respond

less to electricity costs, and vice versa. Richer households of a smaller size, where the head is older

and educated respond more to electricity costs. This suggests that these households might be the

most sophisticated–enough to rely on the EnergyGuide and more detailed energy information. The

structural model, presented next, explicitly tests this hypothesis, and yields results consistent with

these patterns.

6. Estimation: Information Acquisition Model

To obtain an econometric specification of the information acquisition model, I maintain the as-

sumption that the idiosyncratic taste parameters (ε) are extreme value distributed. In addition, I

21

assume that the information acquisition costs have an unobservable idiosyncratic component, νie,

that is also Type I extreme value distributed. For a level of effort, e, the cost for consumer i is

given by:

(5) Ki(e) = Ke + βeXi + νie,

where Xi is a vector of demographics, and the constant Ke and the vector βe are coefficients to be

estimated.

The choice model takes the following general form:

(6) Qirt(j) =∑

e={U,ES,I}Hirt(e)P eirt(j),

where Hirt(e) are the effort choice probabilities, and P eirt(j) are the product choice probabilities

conditional on e. The alternative-specific utilities, Uijrt, that enter P eirt(j) for each level of effort

are:

e=I: U Iijrt = −ηPjrt + ψRrt ×Djt + τ IDjt − θCjr + δj + εIijrt

e=ES: UESijrt = −ηPjrt + ψRrt ×Djt + τESDjt − θESAV INGSrt ×Djt + δj + εESijrt

e=U: UUijrt = −ηPjrt + γj + εUijrt.

When e = I, consumers form a perfect forecast of the operating costs for each product in the

choice set by multiplying the annual electricity consumption of refrigerator j and the average county

electricity price in region r. When e = ES, the term ESAV INGSrt is the difference between the

average operating costs of ES and non-ES refrigerators, in region r at time t. The effect of the label

is included for both e = I and e = ES, which means that consumers can value certified models

beyond purely energy savings in both cases. The coefficient can, however, vary across types. The

effect of rebates is also included for both e = I and e = ES. If consumers are uninformed (e = U),

they will trade-off refrigerators only with respect to prices and non-energy related attributes.

The effort choice probabilities are given by:

(7) Hirt(e) = exp (−Ke − βeXi + Eε,D,S,C [maxj{Uijrt}|I(e)])∑k exp (−Kk − βkXi + Eε,D,S,C [maxj{Uijrt}|I(k)]) .

22

Evaluating the expectation in (7) has two challenges: it requires specifying consumers’ beliefs

about electricity costs and ES, and then integrating over a multi-dimensional integral. The latter

is particularly challenging for large choice sets, as in the present application (Table 1). The for-

mer brings an identification issue given that consumers’ beliefs are not observed at the moment

they make a purchase decision. This implies that the information acquisition costs that enter the

effort choice probabilities are not identified given that their magnitude depend on how beliefs are

specified. I circumvent these two difficulties by using a flexible representation of the expectation at

different levels of effort. In particular, I approximate the expectation with variables that capture

the characteristics of the choice set faced by each consumer, and related to the decision to process

or not energy information.

As discussed by Sallee (2013), in a rational search model consumers might be inattentive to

energy efficiency if the variance in the value of other attributes is large relatively to the variance in

electricity costs. The intuition is that the ex ante value of information increases with the variance

in beliefs. Therefore, if consumers have limited resources to spend on processing information, they

should allocate attention such that learning about attributes has the highest expected returns.

With this intuition in mind, the effort choice probabilities are specified as follows:

(8) Hirt(e) = eVirt(e)∑k e

Virt(k) ,

where

Virt(e = I) =−KI − βIXi + γI1MeanElecrt + γI2V arElecrt + γI3NbModelsrt(9)

+ γI4V arPricert + γI5V arFErt,

Virt(e = ES) =−KES − βESXi + γES1 MeanESrt + γES2 V arESrt + γES3 NbModelsrt

+ γES4 V arPricert + γES5 V arFErt,

Virt(e = U) = 0.

The variables MeanElecrt and V arElecrt are the mean and variance in electricity costs for all prod-

ucts offered in region r at time t, MeanESrt is the proportion of ES models offered, NbModelsrt is

the number of models in the choice set in a given region, and V arPricert is the variance in prices.

23

Finally, V arFErt represents the variance in quality offered in each store, which is captured by the

product fixed effects.

The estimation is carried with a two-step maximum likelihood procedure. In the first step,

the product fixed effects are recovered by estimating a simple conditional logit with no unobserved

heterogeneity. The second step estimates the information acquisition model using the product fixed

effects from the conditional logit as data. The standard errors in the second step are corrected using

the approach of Murphy and Topel (1985). To capture heterogeneity with respect to income, the

model is estimated separately for the three different income groups.

6.1. Identification

The information acquisition model is a special case of a mixed logit with three discrete latent

classes. It can also be interpreted as a sorting model, where the effort choice probabilities capture

why consumers self-select to attend to a particular piece of energy information. The information

acquisition costs are not identified because consumer beliefs are not specified. The goal of the

estimation is to identify the share of consumers that fall in each latent class, and the behavioral

parameters that enter the choice probabilities (η, θ, τ I , τES , and ψ). The sources of variation

used to identify the coefficients on prices, electricity costs, rebates, and electricity costs have been

discussed in Sections 4 and 5. In what follows, I discuss identification of the latent classes.

The latent classes capture unobserved heterogeneity in the valuation of energy efficiency. As it

is customary in discrete choice models, events that induce specific substitution patterns identify

heterogeneity. In the present application, the coarseness of the certification requirement versus the

continuous nature of the information related to electricity costs each yields different substitution

patterns, which allows us to classify consumers. To illustrate, consider a market with only three re-

frigerator models, all with the same quality, but with different levels of energy consumption. Figure

6 represents the location of each refrigerator model in the quality-energy efficiency characteristics

space, before and after the revision in the ES standards, for three types of consumers. Panels (a)

and (d) represent the location of the refrigerator models in the quality-energy cost characteristics

space for consumers that are fully informed about energy costs. For this consumer type, products

are located at a different address in the characteristics space. How market shares will vary for a

24

relative change in prices will then be a function of the distance between products due to the dif-

ference in electricity costs. Note that for this consumer type, a revision in the ES standard (Panel

(d)), represented by the line s, should not influence the substitution patterns. Panels (b) and (e)

represent the location of the same models in the quality-ES characteristics space. This corresponds

to a graphical representation of consumers that rely on the ES certification to account for the

energy efficiency attribute. These consumers perceived ES refrigerators as perfect substitutes and

the revision in the ES standard impacts this perception (Panel (e)). For this consumer type, the

Independence of the Irrelevant Alternative (IIA) hypothesis should hold within the class of products

that are certified (and with similar quality). Finally, for consumers that do not account for the

energy efficiency attribute at all (panels (c) and (f)), all products are located at the same address in

the characteristics space and are perceived as perfect substitutes. The IIA hypothesis should hold

for all products with similar quality, irrespective of their location in the energy efficiency dimension.

To summarize, the existence of consumers with different degrees of sophistication with respect

to the way they account for energy efficiency implies different latent substitution patterns. The

discrete and binary nature of the ES certification allows us to distinguish consumers that trade off

energy efficiency using the certification from consumers that use a more continuous measure, such

as an estimate of electricity costs provided by EnergyGuide. Ultimately, each consumer type is

identified by events that cause substitution between products. In the data, relative price changes

and product entry and exit are two such types of events. They are both frequent and important.

The decertification event, as illustrated above, also contributes to the identification.

6.2. Results

Table 5 presents the estimation results for each income group using the sub-sample for the year 2008.

Comparing the estimates of the coefficient on electricity costs (θ) of the information acquisition

model with the conditional logit (Table 3) shows the importance of controlling for heterogeneity

in consumer sophistication. While the conditional logit model suggests a large undervaluation of

energy efficiency, this phenomenon has a more nuanced interpretation in the information acquisition

model. For consumers that rely on electricity costs (e = I), the implicit discount rate is low and on

par with other investment decisions. Therefore, the a priori hypothesis that a share of consumers

25

behave as if they were perfectly informed holds true. In the present framework, the existence of a

so-called Energy Efficiency Gap (Jaffe and Stavins 1994) is attributable to the existence of a share

of uninformed consumers that dismiss the energy efficiency attribute altogether. This share is 65%,

41%, and 33% for the low, medium, and high income group, respectively.

With the information acquisition model, the issue of measurement error in average electricity

prices also becomes a moot point. The share of perfectly informed consumers corresponds to

consumers that are sophisticated enough to use county average electricity prices in their decision.

As long as this is the share of consumers that I am interested to identify for the policy analysis,

county averages are the appropriate level of aggregation. If I were to use a different measure of

electricity prices, such as state averages, the share of perfectly informed consumers would simply

correspond to a different type of consumers.

The share of consumers that rely primarily on ES (e = ES) increases with income. These

consumers have a high WTP for the label itself (coefficient τm), which ranges from $95 to $152.

This implies that some consumers value the certification well beyond the average energy savings

associated with the certification. On average, the discount value of the energy savings associated

with ES are $83 for the year 2008.12

On the other hand, consumers that are fully informed (e = I) have a low WTP for ES, and even

negative for the lower income group. This suggests that the ES certification and information about

electricity costs are substitute. For some consumers, the ES label is simply superfluous, but for

others, it is a highly valued attribute. Whether ES is a useful heuristic from a welfare standpoint

is, however, unclear. The large willingness to pay for the label itself is puzzling, and suggests that

consumers might have a biased perception of the overall quality of certified products.

Relative to the conditional logit (Table 3), the coefficients on rebate increase, are positive, and

become significant for the lower income groups. Again, this shows the importance of accounting

for heterogeneity in sophistication. Lower income households have a higher propensity to take

advantage of a rebate program, which is consistent with the idea that, relative to richer households,

they may be more willing to overcome the hassle costs of claiming a rebate.12Assuming that non-ES refrigerators consume 568 kWh/year and ES refrigerators consume 500 kWH/year,a lifetime of 18 years, and an average electricity price of 0.113 $/kWh (Table 1).

26

The coefficients that enter the latent probabilities are highly consistent with the results of the

conditional logit interacted with demographics (Table 4). Again, I find that richer, but smaller

households, where the head is older and has a graduate degree are more likely to rely on electricity

costs. Age and education are also correlated positively with the probability of relying on ES.

Republicans, relative to Democrats or households that have other political orientation, are more

likely to rely on the certification. The variables that capture the choice set faced by each consumer

tend to not have a significant impact. The higher the proportion of ES models offered, the more

likely consumers are to rely on the certification.

7. Alternative Estimator: Heterogeneity

The information acquisition model introduces heterogeneity along observable and unobservable di-

mensions in a parsimonious, but restrictive manner. This section presents a more flexible estimator

to capture unobserved heterogeneity. In particular, I recover the nonparametric joint distribution

of the coefficients on the ES label (τ) and electricity costs (θ) using the estimator proposed by

Fox et al. (2011) (FKRB). The estimator provides one more piece of evidence that the preference

parameters τ and θ are inversely correlated. It also shows that the information acquisition model

captures well the main patterns of unobserved heterogeneity.

The estimator works as follows. Suppose that τ has support on the [τ , τ ] interval, and similarly

θ has support on [θ, θ]. FKRB propose to take a large but finite number of grid points on the

intervals, and treat each point as a random coefficient with an unknown frequency that needs to

be estimated. Taking all combinations of grid points, say M , I can form the choice probabilities

for each combination, m, and approximate the choice probability for each consumer i by averaging

over all the random coefficients:

(10) Qijrt ≈M∑m=1

αm · Pmijrt(βm),

where αm is the unknown frequency of the pair of random coefficients: βm = {τm, θm}, and

Pmijrt(βm) is the choice probability evaluated at βm. As shown by FKRB, the fact that the dependent

27

variable of the model Pijrt is linearly related to the parameters αm allows us to consistently estimate

αm using inequality constrained least squares.13

For the present application, I fix η, ψ and γj at their MLE estimates (Model 1, Table 2), and

center the support of τ and θ around the MLE estimates.14 To recover a smooth joint distribution

of random parameters, I model the joint density of τ and θ as a mixture of normal basis densities

instead of point masses, as suggested by FKRB. Using the simpler estimator with point masses

produces qualitatively similar results. Additional details on the implementation can be found in

Appendix D.

Figure 5 plots the contour map of the estimated joint density of the parameters θ and τ for

different income groups. For lower income households (Figure 5(a)), two clusters are identified,

one with a relatively low valuation to energy costs and ES, and a second with a low sensitivity to

energy costs, but a high valuation of ES. For households in the second income group (Figure 5(b)),

the distribution has a similar pattern. For higher income households (Figure 5(c)), three large

clusters are identified: households with a low valuation of ES and electricity costs; a low valuation

of ES, but a high valuation of electricity costs, and a high valuation of ES and low valuation of

electricity costs. A fourth, but marginal cluster, where households have a high valuation of both

ES and electricity costs is also detected.

Altogether, the heterogeneity patterns identified are consistent with the fact that preferences

for the ES certification and a continuous measure of electricity costs are inversely correlated. Con-

sumers that rely on a particular piece of information are less likely to rely on the other, and vice

versa. Moreover, the FKRB estimator also reveals that a discrete model with few latent classes,

as in the information acquisition model, is sufficient to capture the main sources of unobserved

heterogeneity. It also suggests that using a mixed logit model relying on an unimodal distribution

to capture unobserved heterogeneity, as it is customary in the literature, will provide a poor fit to

the data.13This estimator is thus appealing because of its computational simplicity. It is, however, prone to sufferfrom the curse of dimensionality, and requires some subjectivity in the choice of grid points.14The estimator is a special case of a two-step estimator. FKRB propose a bootstrap procedure to recoverthe standard errors in the second step. An approach similar to Murphy and Topel (1985) can also be usedto account for the fact that some parameters that enter the choice model were estimated in the first step. Inthe present application, the estimated joint distributions are presented graphically, without standard errors.

28

8. The Value of Energy Information

This section shows how consumers, firms, and society as a whole might lose, or even gain, when some

consumers rely on ES or dismiss energy efficiency altogether instead of being perfectly informed

about energy costs. In particular, I quantify the opportunity costs of imperfect energy information

in the US refrigerator market. This is a first step toward performing a welfare analysis of ES. A

complete analysis would require to compare these costs to the economies in the costs of collecting

and processing information brought by the certification. I do not quantify these economies. The

concept of opportunity cost of imperfect energy information is in itself important for policy design.

It provides an upper bound on the net benefits of better informing consumers in this market.

Formally, I define the opportunity cost of imperfect energy information as the difference between

the welfare that would be achieved if all consumers were to make a purchase perfectly informed,

versus the expected welfare that is achieved when consumers are relying on ES or dismissing energy

efficiency. The structural model can readily be used to simulate these quantities. If all consumers

are perfectly informed, the choice probabilities are given by: Qirt(j) = P e=Iirt (j), and the expected

welfare is:

(11) E [Wirt|I(e = I)] = 1J

J∑j

P e=Iirt (j) · (CSj + PSj − EXTj) ,

where CSj is a measure of consumer surplus associated with product j, PSj is the producer surplus,

and EXTj are the externality costs. The opportunity costs of relying on ES or being uninformed

are, respectively, given by:

E [Wirt|I(e = I)]− E [Wirt|I(e = ES)] =(12)

1J

J∑j

(P e=Iirt (j)− P e=ESirt (j)

)· (CSj + PSj − EXTj),

E [Wirt|I(e = I)]− E [Wirt|I(e = U)] =

1J

J∑j

(P e=Iirt (j)− P e=Uirt (j)

)· (CSj + PSj − EXTj),

where P e=ESirt (j) and P e=Uirt (j) correspond to the choice probabilities of each consumer type.

29

Consumer Surplus. According to the information acquisition model, for consumers that do not

fully process energy information, there is a discrepancy between the electricity costs they believe

they would pay and the electricity costs they effectively pay. That is, the utility they experience

differs from the utility they thought they would experience. This gap between decision and expe-

rience utility raises the question of what concept of utility should be used to measure consumer

welfare. I propose a measure of consumer surplus based on the notion of experience utility, and

make the following assumption.

Assumption 1. If e = I, decision utility equals experience utility.

Assumption 1 simply says that under perfect information consumers experience what they be-

lieved they would experience. Under this assumption, the consumer surplus is the observed com-

ponent of experience utility corresponding to the perfectly informed consumers:

(13) CSijrt = 1η

(γj + τDjt − ηPjrt + ψRrt ×Djt − θCjr) .

In particular, I have divided by the marginal utility of income to convert utils into dollars.15

Whether the effect of the ES label is truly experienced, and whether τ I or τES should be

used can be debated. A key insight from the empirical analysis is that consumers that rely on

ES value certified products well beyond purely energy savings. Whether this high willingness

to pay for the label itself reflects a behavioral bias or corresponds to preferences has important

welfare implications. If consumers are subject to a bias akin of the halo effect, i.e., wrongfully

believe that certified products are of higher quality simply because of the label, they will pay too

much for certified products. Under this interpretation, the label effect might then reduce welfare.

Empirically, τ I tend to be close to zero, and τES to be large. I then propose to bound the change

in consumer surplus by considering a measure that simply excludes the label effect, and a measure

that includes τES .

Producer Surplus. For each refrigerator model in the sample, I observe the wholesale price. I

define the producer surplus as the difference between the retail and wholesale price. This measure15Note that the unobserved idiosyncratic taste parameters εijrt do not appear in the consumer surplus. Themeasure thus corresponds to the expected consumer surplus, because I use the fact that Eε[εijrt] = 0. Also,note that (13) should be interpreted as the expected consumer surplus over the lifetime of the refrigerator.

30

of producer surplus should be interpreted as a lower bound, because it excludes manufacturers’

profits.16

Externality Costs. To quantify the externality costs associated with the electricity generated

to operate refrigerators, I focus on the emissions of greenhouse gases (CO2, CH4, N2O), sulphur

dioxide (SO2), and nitrous oxide (NOx), and use emissions factors recommended by the EPA. The

externality costs associated with each scenario are the product of the corresponding average elec-

tricity consumption purchased, the emission factors, and the damage costs of electricity generation,

where the average electricity consumption purchased is the average of the electricity consumption

of the refrigerators sold, weighted by market shares.

The dollar damages associated with carbon dioxide correspond to the recent estimates of the

social cost of carbon recommended to assess federal regulations (Greenstone, Kopits and Wolverton

2011). For sulphur dioxide and nitrous oxide, I rely on the estimates used by the Department of

Energy in the cost-benefit analysis of the 2014 minimum energy efficiency standards for refrigerators

(DOE 2011), and the estimates provided by Muller and Mendhelson (2012). Table 9 presents the

emission factors and the damage costs. The results that I report use the high estimates.

8.1. Results

ENERGY STAR at 20%. In this scenario, the stringency of the ES requirement is set at 20%

below the federal minimum energy efficiency standard.17 The demand model is simulated with

the estimates of Table 5 using a random sample of 2,000 households. The electricity prices, and

rebates faced by each household are the same as in the demand estimation. The products on the

market are all refrigerator models observed in the year 2008 that were included in the demand

estimation. Figure 7(a) shows the distribution of the products with respect to energy efficiency.

Products bunch at the minimum and ES standards. There is also a large number of products that

meet the previous certification requirement (15% more efficient than the minimum). The price of

each model is the average price observed in the post-revision period (May-December 2008).16Under most forms of competition between upstream and downstream firms, I conjecture that the changein profits between manufacturers and retailers are of the same sign (Bresnahan and Reiss 1985).17A product can be certified if it consumes at least 20% less electricity than a similar model that just meetthe minimum standard.

31

Table 6 presents the estimates of the opportunity costs with respect to the various metrics that

enter welfare. The first column compares the outcomes obtained under the information acquisition

model to a model where all consumers make a purchase decision perfectly informed. These estimates

correspond to the opportunity costs of imperfect energy information under the current labeling

policies. The opportunity costs of being completely uninformed or relying solely on ES are presented

in the second and third column, respectively.

Focusing on the measure of consumer surplus that includes the label effect (τES), consumers lose

$19, on average, under the information acquisition model. This estimate can be interpreted as an

upper bound on the additional costs of collecting and processing energy information that consumers

that rely on ES or dismiss energy efficiency should be ready to invest to become perfectly informed.

Put simply, this estimates suggest that consumers are willing to forgo $19, on average, to not have to

deal with energy information. The opportunity cost of being completely uninformed about energy

efficiency is larger: $28 on average for all income groups, and is the largest for the lowest income

households ($33). If all consumers were to rely on ES, the opportunity costs would be $21. This is

an improvement relative to the case where all consumers are uninformed. This conclusion, however,

only holds if the label effect is included in the consumer surplus. If excluded, the opportunity cost

of relying on ES increases to $42, while the opportunity cost of being uninformed is only $29. Here,

although ES provides some information about energy efficiency, it still makes consumers worst-off,

relative to the scenario where no information is provided. This is a counterintuitive result. This

shows that by excluding the label effect from the consumer surplus, the certification is effectively

acting as a bias. Ex post, consumers are then worst-off because the label led them to pay too much

for certified products.

Turning to other metrics, under perfect information, consumers choose more energy efficient

refrigerators, on average. The reduction in electricity consumption under ES is similar to the

reduction achieved with only uninformed consumers. The certification then brings modest welfare

gains from the reduction in externality costs. The fact that the certification requirement is set in

relative terms and depends on other attributes, such as size, explains this result.

The size of the externality costs under the information acquisition model is $14 per refrigerator

sold. The opportunity cost of imperfect energy information measured in terms of consumer surplus

32

net of the externality costs thus ranges between $34 to $40, depending on how the label effect is

interpreted.

Firms are better off when consumers are not perfectly informed, and would gain the most if all

consumers were to rely on ES. This implies that firms maintain larger markups on certified products.

This estimate should, however, be interpreted with an important caveat. If all consumers were

perfectly informed, firms would make different product line and pricing decisions. In the medium-

term, the impact on profits could then be different. In a related paper (Houde 2013), I show that

once product line and pricing decisions are endogenized, the large willingness to pay for the ES

label allows firms to charge larger markups not only on certified products, but also on non-certified

products. In sum, firms appear to benefit largely from the existence of the certification both in the

short and long-run.

The above results also suggest that firms’ provision of energy information may not be fully

compatible with the social planner’s objective. Using a similar framework, Sallee (2013) shows that

firms may be better off distorting the provision of other non-energy attributes so that consumers

pay less attention to energy efficiency.

Overall, once I account for the change in consumer surplus, externalities, and profits, there

is an opportunity cost associated with consumers not being perfectly informed that ranges from

$12 to $17 per refrigerator sold. Assuming a market size of 9.01 million (DOE 2011), the value

of perfect energy information is more than $108M/year for the refrigerator market alone. To put

this number in perspective, the EPA and the DOE have spent $57.4 M/year, on average, during

the period 2008-2011 to run the whole ES program (US Government Accountability Office GAO),

which includes more than sixty categories of products.

8.2. Sensitivity Tests

In the online Appendix E, several sensitivity tests are presented. The results are summarized below.

Electricity Prices, Rebates. Electricity prices and financial incentives for ES products vary

widely across the US. Table 10 provides estimates of the opportunity costs for a scenario with low

33

electricity prices, and no rebate. In this scenario, the average electricity price is set at 0.08 $/kWh

for all consumers.

As expected, lower electricity prices translate in a lower value of energy information for con-

sumers. Consumers are simply more likely to not pay attention to energy efficiency. In this

scenario, the overall opportunity cost, accounting for the change in consumer surplus, externalities,

and profits is between $4 to $8.

Table 11 presents estimates where the electricity price is set at 0.16 $/kWh, and a $50 rebate

is offered to all consumers. In this scenario, the value of energy information is much larger for

consumers, and the opposite for firms. On the other hand, the externality costs remain unaffected,

relative to the previous scenarios. The overall opportunity cost now ranges from $21 to $25.

Choice Set. In constructing the choice set for each household, I use all refrigerator models present

in the sample during the year 2008. As a result, the choice sets include several models that meet the

previous certification requirement, but are no longer certified. As observed on Figure 7(b), these

products exit the market in the long term. In equilibrium, firms try to maintain a distribution of

products that either just meet the minimum standard or the ES certification. In this scenario, I

modified the choice sets to reflect such an equilibrium. In particular, all models that just meet the

previous requirement, i.e., that are 15% more efficient than the minimum standard, were excluded

(Figure 7(b)).

The main take away is that under this scenario, relying solely on ES has a much lower opportunity

cost (Table 12). If there is close to perfect bunching at the minimum standard and ES in the product

offering, the loss in accuracy incurred by trading-off energy efficiency in a binary manner becomes

minimal. This explains that bunching can be sustained in equilibrium. If firms offer products that

bunch at the minimum standard and ES, consumers have higher incentive to use the certification as

a decision heuristic, which in turn reinforces firms’ incentives to bunch. Whether these distortions

are welfare improving is unclear. The coarse nature of the certification crowds out consumers’ and

firms’ incentives to purchase and offer products that exceed the ES requirement. As a result, the

certification might have the unintended consequence to reduce the provision of energy efficiency.

34

9. Conclusion

This paper first shows that consumers, on average, respond to both the ES certification and elec-

tricity costs in the appliance purchasing decision. It then provides evidence that consumers are

heterogeneous in the way they respond to different pieces of energy information. The main contri-

bution of the paper is a structural estimator that classifies consumers in three types: consumers

that rely on a continuous measure of electricity costs, consumers that rely on the ES certification,

and consumers that do not pay attention to energy-related information. Identification comes from

the fact that consumer types respond differently to changes in relative prices, choice sets, and de-

certification events; each type is thus associated with a specific substitution pattern. Empirically,

the prevalence of each substitution pattern determines the share of consumers that belongs to each

type.

The estimator provides several insights on how a certification that summarizes complex, but

easily accessible information influences choices. In the present context, for the share of consumers

that rely on an accurate measure of electricity costs, the certification appears to have no effect. On

the other hand, consumers that rely on ES have a high willingness to pay for the label itself that

is well beyond the value of the energy savings associated with certified products. From a welfare

standpoint, the interpretation of the label effect has important implications. Further research is,

however, needed to better identify the mechanisms that explain the strong effect of the ES label,

and whether those mechanisms apply to other contexts.

The policy analysis shows that a consumer in the US refrigerator market loses between $19 to

$24, on average, for not being perfectly informed about electricity costs. If all consumers were

perfectly informed, choices would be more energy efficient, and the externality costs would be

reduced by $14 per refrigerator sold. Interestingly, the opportunity costs measured in terms of

externality costs do not differ by a large amount whether consumers rely solely on the certification

or are completely uniformed about electricity costs. Firms are better off when not all consumers

are perfectly informed, and benefit the most when consumers rely on ES.

Accounting for the change in consumer surplus, externality costs, and profits, I find that the

overall opportunity cost of imperfect energy information ranges from $12 to $17 per refrigerator

35

sold. Scaling these estimates by the size of the refrigerator market, the total size of the opportunity

cost is $108M/year, a number twice as large the annual program costs of ES.

References

Abaluck, Jason. 2011. “What Would We Eat if We Knew More: The Implications of a Large-Scale

Change in Nutrition Labeling.” Massachusetts Institute of Technology Working Paper.

Abaluck, Jason, and Jonathan Gruber. 2011. “Choice Inconsistencies among the Elderly:

Evidence from Plan Choice in the Medicare Part D Program.” The American Economic Review,

101(4): 1180–1210.

Akerlof, George A. 1970. “The Market for ”Lemons”: Quality Uncertainty and the Market

Mechanism.” The Quarterly Journal of Economics, 84(3): pp. 488–500.

Aldy, Joseph, and Sebastien Houde. 2014. “Belt and Suspenders and More: The Incremental

Impact of Energy Efficiency Subsidies in the Presence of Existing Policy Instruments.” Working

Paper.

Allcott, Hunt. 2011. “Consumers’ Perceptions and Misperceptions of Energy Costs.” American

Economic Review, 101(3): 98–104.

Allcott, Hunt, and Nathan Wozny. 2012. “Gasoline Prices, Fuel Economy, and the Energy

Paradox.” National Bureau of Economic Research Working Paper 18583.

Anderson, Soren T, Ryan Kellogg, James M Sallee, and Richard T Curtin. 2011. “Fore-

casting gasoline prices using consumer surveys.” American Economic Review, 101(3): 110.

Bagwell, Kyle. 2007. “Chapter 28 The Economic Analysis of Advertising.” In . Vol. 3 of Handbook

of Industrial Organization, , ed. M. Armstrong and R. Porter, 1701 – 1844. Elsevier.

Bento, Antonio M., Shanjun Li, and Kevin Roth. 2012. “Is there an energy paradox in fuel

economy? A note on the role of consumer heterogeneity and sorting bias.” Economics Letters,

115(1): 44 – 48.

Berry, Steven T. 1994. “Estimating Discrete-Choice Models of Product Differentiation.” The

RAND Journal of Economics, 25(2): pp. 242–262.

36

Blumstein, C., B. Krieg, L. Schipper, and C. York. 1980. “Overcoming Social and Institu-

tional Barriers to Energy Conservation.” Energy, 5: pp. 355–371.

Borenstein, Severin. 2010. “The Redistributional Impact of Non-Linear Electricity Pricing.”

Energy Institute at Haas Working Paper 204.

Bresnahan, Timothy F., and Peter C. Reiss. 1985. “Dealer and Manufacturer Margins.” The

RAND Journal of Economics, 16(2): pp. 253–268.

Busse, Meghan R., Christopher R. Knittel, and Florian Zettelmeyer. 2013. “Are Con-

sumers Myopic? Evidence from New and Used Car Purchases.” American Economic Review,

103(1): 220–56.

DOE. 2011. “Energy Conservation Program: Energy Conservation Standards for Residential Re-

frigerators, Refrigerator-Freezers, and Freezers.” Federal Register, 76(179).

Dranove, David, and Ginger Zhe Jin. 2010. “Quality Disclosure and Certification: Theory

and Practice.” Journal of Economic Literature, 48(4): 935–63.

EIA. 2008. “Form EIA-861.” Energy Information Administration

http://www.eia.gov/cneaf/electricity/page/eia861.html.

Fox, Jeremy T., Kyoo Kim, Stephen P. Ryan, and Patrick Bajari. 2011. “A Simple

Estimator For The Distribution Of Random Coefficients.” Quantitative Economics, 2(3): 381 –

418.

Gabaix, Xavier, David Laibson, Guillermo Moloche, and Stephen Weinberg. 2006.

“Costly Information Acquisition: Experimental Analysis of a Boundedly Rational Model.” Amer-

ican Economic Review, 96(4): 1043–1068.

Greenstone, Michael, Elizabeth Kopits, and Ann Wolverton. 2011. “Estimating the Social

Cost of Carbon for Use in U.S. Federal Rulemakings: A Summary and Interpretation.” National

Bureau of Economic Research, Inc NBER Working Papers 16913.

Handel, Benjamin R., and Jonathan T. Kolstad. 2013. “Health Insurance for ŞHumansŤ:

Information Frictions, Plan Choice, and Consumer Welfare.” National Bureau of Economic Re-

search Working Paper 19373.

37

Houde, Sebastien. 2013. “Bunching with the Stars: How Firms Respond to Environmental Cer-

tification.” University of Maryland Working Paper.

Imbens, Guido W., and Joshua D. Angrist. 1994. “Identification and Estimation of Local

Average Treatment Effects.” Econometrica, 62(2): pp. 467–475.

IREC. 2011. “Database for State Incentives for Renewable and Efficiency.”

http://www.dsireusa.org/.

Ito, Koichiro. 2012. “Do Consumers Respond to Marginal or Average Price? Evidence from

Nonlinear Electricity Pricing.” National Bureau of Economic Research Working Paper 18533.

Jaffe, Adam B, and Robert N Stavins. 1994. “The energy-efficiency gap What does it mean?”

Energy policy, 22(10): 804–810.

Klier, Thomas, and Joshua Linn. 2010. “The Price of Gasoline and New Vehicle Fuel Economy:

Evidence from Monthly Sales Data.” American Economic Journal: Economic Policy, 2(3): 134–

53.

Li, Shanjun, Christopher Timmins, and Roger H. von Haefen. 2009. “How Do Gasoline

Prices Affect Fleet Fuel Economy?” American Economic Journal: Economic Policy, 1(2): pp.

113–137.

Mason, Charles F. 2013. “The Economics of Eco-Labeling: Theory and Empirical Implications.”

Center for Energy and Environmental Economics Working Paper.

McCall, John J. 1965. “The Economics of Information and Optimal Stopping Rules.” The Journal

of Business, 38(3): pp. 300–317.

McWhinney, Marla, Andrew Fanara, Robin Clark, Craig Hershberg, Rachel Schmeltz,

and Judy Roberson. 2005. “ENERGY STAR Product Specification Development Framework:

Using Data and Analysis to Make Program Decisions.” Energy Policy, 33(12): 1613 – 1625.

Meier, Alan K., and Jack Whittier. 1983. “Consumer Discount Rates Implied by Purchases

of Energy-Efficient Refrigerators.” Energy, 8(12): 957 – 962.

Muller, Nicholas Z., and Robert Mendelsohn. 2012. “Efficient Pollution Regulation: Getting

the Prices Right: Corrigendum (Mortality Rate Update).” American Economic Review, 102(1).

38

Newell, Richard G., and Juha Siikamaki. 2013. “Nudging Energy Efficiency Behavior, The

Role of Information Labels.” Resources for the Future RFF Discussion Paper 13-17.

Sallee, James M. 2013. “Rational Inattention and Energy Efficiency.” National Bureau of Eco-

nomic Research Working Paper 19545.

Sims, Christopher A. 2003. “Implications of Rational Inattention.” Journal of Monetary Eco-

nomics, 50(3): 665 – 690.

Stigler, George J. 1961. “The Economics of Information.” Journal of Political Economy, 69(3): pp.

213–225.

US Government Accountability Office (GAO). 2011. “ENERGY STAR Providing Oppor-

tunities for Additional Review of EPA’s Decisions Could Strengthen the Program.” Report to

Congressional Requesters. GAO-11-888.

Viscusi, W Kip. 1978. “A Note on ”Lemons” Markets with Quality Certification.” Bell Journal

of Economics, 9(1): 277–279.

Ward, David O., Christopher D. Clark, Kimberly L. Jensen, Steven T. Yen, and

Clifford S. Russell. 2011. “Factors influencing willingness-to-pay for the ENERGY STAR

label.” Energy Policy, 39(3): 1450 – 1458.

39

10. Figures

(a) ES label (b) EnergyGuide label

Figure 1. Energy Labels for the US Appliance Market

40

400

450

500

550

600

Ele

ctric

ity (

kwh/

year

)

15 20 25 30 35

2006

400

450

500

550

600

15 20 25 30 35

2007

400

450

500

550

600

15 20 25 30 35

2008

400

450

500

550

600

15 20 25 30 35

2009

400

450

500

550

600

10 15 20 25 30 35

Adjusted Volume (Cu.Ft.)

2010

Minimum Energy Efficiency StandardENERGY STAR: 2001ENERGY STAR: 2004ENERGY STAR: 2008

Minimum Energy Efficiency Standard

ENERGY STAR: Pre-2008 Revision

ENERGY STAR: Post-2008 Revision

400

450

500

550

600

15 20 25 30 35

Standards

(a) Bottom-Freezer Refrigerator without Ice-Maker

400

450

500

550

600

650

700

750

800

Ele

ctric

ity (

kwh/

year

)

15 20 25 30 35 40

2006

400

450

500

550

600

650

700

750

800

15 20 25 30 35 40

2007

400

450

500

550

600

650

700

750

800

15 20 25 30 35 40

2008

400

450

500

550

600

650

700

750

800

15 20 25 30 35 40

2009

400

450

500

550

600

650

700

750

800

15 20 25 30 35 40Adjusted Volume (Cu.Ft.)

2010

Minimum Energy Efficiency Standard

ENERGY STAR: Pre−2008 Revision

ENERGY STAR: Post−2008 Revision

400

450

500

550

600

650

700

750

800

15 20 25 30 35

Standards

(b) Side-by-Side Refrigerator with Ice-Maker

400

450

500

550

600

Ele

ctric

ity (

kwh/

year

)

10 15 20 25 30 35

2006

400

450

500

550

600

10 15 20 25 30 35

2007

400

450

500

550

600

10 15 20 25 30 35

2008

400

450

500

550

600

10 15 20 25 30 35

2009

400

450

500

550

600

10 15 20 25 30 35Adjusted Volume (Cu.Ft.)

2010

Minimum Energy Efficiency StandardENERGY STAR: 2001ENERGY STAR: 2004ENERGY STAR: 2008

Minimum Energy Efficiency Standard

ENERGY STAR: Pre-2008 Revision

ENERGY STAR: Post-2008 Revision

400

450

500

550

600

15 20 25 30 35

Standards

(c) Top-Freezer Refrigerator without Ice-Maker

Figure 2. Choice Set Full-Size Refrigerators: 2006-2010

41−

100

010

020

030

040

050

0%

Cha

nge

in W

eekl

y S

ales

w.r

.t A

vg W

eekl

y S

ales

−40 −20 0 20% Change in Retail Price w.r.t Suggested Retail Price

2008

−10

00

100

200

300

400

500

% C

hang

e in

Wee

kly

Sal

es w

.r.t

Avg

Wee

kly

Sal

es

−40 −20 0 20% Change in Retail Price w.r.t Suggested Retail Price

2010

Figure 3. Correlation Between Sales and Prices

.9.9

51

1.05

1.1

Nor

mal

ized

Pric

e

2008w1 2008w14 2008w27 2008w40 2009w1Weeks, Year 2008

Non Energy Star Models

Decertified Energy Star Models: 15% Better than Minimum

Energy Star Models: 20% Better than Minimum

2008

.85

.9.9

51

1.05

1.1

Nor

mal

ized

Pric

e

2009w13 2009w40 2010w13 2010w40Weeks, Year 2009−2010

Energy Star Models

Decertified Energy Star Models

Non Energy Star Models

2009−2010

Figure 4. Normalized Prices Before and After Decertification

Each panel displays average normalized weekly prices, with 5% confidence intervals, ofrefrigerators that belong to different efficiency classes. The left panel shows that the priceof decertified models slightly decrease after the decertification, but the change is small. Theright panel shows a large, but temporary price decrease following the decertification event.

42

(a) Income < $50, 000

(b) Income ≥ $50, 000, < $100, 000

(c) Income ≥ $100, 000

Figure 5. Joint Density of τ and θ.

Colored areas correspond to the regions with a positive density above a small cut-off value(1e-5). Red areas correspond to the regions with the highest density.

43

(a) (b) (c)

(d) (e) (f)

Figure 6. (a) Consumers with expectations about energy costs for eachmodel. s represents the ES standard. (b) Consumers that rely on ES. D = 1models certified ES, zero otherwise. (c) Consumers that Dismiss Energy Costs.(d)-(f) Change in consumer preferences after revision of the ES standards→ s′.

44

0.1

.2.3

Den

sity

0 5 10 15 20 25 30Energy Efficiency: Percentage Better than Minimum Standard

(a)

0.1

.2.3

.4D

ensi

ty

0 5 10 15 20 25 30Energy Efficiency: Percentage Better than Minimum Standard

(b)

Figure 7. Choice Sets for Policy Analysis

Each panel corresponds to the density of the energy efficiency levels of all refrigerator modelsoffered. The x-axis is the percentage difference in energy efficiency relative to the federalminimum energy efficiency standards. In the second panel, all products meeting the EScertification requirement prior April 2008 were excluded from the choice set.

45

11. TablesTable 1. Summary Statistics

2008 2010Choice SetNb of Refrigerator Models, US Market 2,524 1,496Nb of Decertified Refrigerator Models, US Market 1,278 21Nb of Refrigerator Models, Sample 1,483 1,174Nb of Decertified Refrigerator Models in Restricted Sample 674 14Average Size of Choice Set (Store-Trimester) 99 92SD Size of Choice Set (Store-Trimester) 44 37

PriceAvg Promotional Price $1,325 $1,411SD Promotional Price Offered $639 $693Avg Promotion 11% 15%Avg Duration of a Promotion 1.4 week 3.1 weekAvg Promotional Price ENERGY STAR $1,426 $1,368Avg Promotional Price Non-ENERGY STAR $1,240 $1,472Avg Promotional Price Decertified Model $1,416 $2,468

RebateNb of Utility Rebate Programs for ENERGY STAR refrigerators 87 133Nb of State Rebate Programs for ENERGY STAR refrigerators 0 44Avg Rebate Offered $48 $104SD Rebate Offered $54 $144

Electricity Costs: US RefrigeratorsAvg Elec. Price (County) 0.113$/kWh 0.113$/kWhMax Elec. Price 0.03$/kWh 0.03$/kWhMin Elec. Price 0.420$/kWh 0.368$/kWhAvg Elec. Consumption 520 kWh/y 506 kWh/yAvg Elec. Consumption ENERGY STAR 500 kWh/y 501 kWh/yAvg Elec. Consumption Decertified ENERGY STAR 520 kWh/y 547 kWh/yAvg Elec. Consumption Non-ENERGY STAR 568 kWh/y 525 kWh/yNotes: The restricted sample consists of all transactions made by homeowners living in singlefamily housing units that bought no more than one refrigerator in any given year. The number offull-size refrigerator models for the whole US was obtained from the US EPA. According to theFTC data, there were 2,693 full-size refrigerators offered on the US market in 2008.

46

Table 2. Conditional Logit

(I) (II) (III) (IV) (V)CL CL CL CL CL

Base Brand-Week State Elec. Price 2009-2011 2009-2010

Price (η) -0.421∗∗∗ -0.402∗∗∗ -0.422∗∗∗ -0.273∗∗ -0.318∗∗∗(0.010) (0.010) (0.010) (0.099) (0.007)

Elect. Cost (θ) -0.681∗∗∗ -0.666∗∗∗ -1.304∗∗∗ -1.279 -1.497∗∗∗(0.167) (0.166) (0.151) (2.203) (0.208)

Rebate (ψ) 0.006 0.005 0.004 0.025∗ 0.046∗∗∗(0.015) 0.015 (0.015) (0.012) (0.014)

ENERGY STAR (τ) 0.080∗∗ 0.101∗∗ 0.099∗∗ 0.245 0.091(0.029) (0.013) (0.013) (1.419) (0.049)

Interpretation

Price Elasticity -5.47 -5.23 -5.49 -3.55 -4.13

WTP ES τ/η ($) 19.0 25.2 23.4 89.6 28.6

Prob. Take Rebate ψ/η 0.01 0.01 0.01 0.09 0.14

Implied Discount Rate 0.62 0.60 0.32 0.21 0.21Product FE Yes Yes Yes Yes YesBrand-Week FE No Yes No No NoAvg Elec Price County County State County CountyNb Obs 201,509 201,509 201,509 184,645 74,300Nb Clusters 922 922 922 913 912Note: Standard errors clustered at the store level in parentheses: † (p < 0.10) ∗(p < 0.05), ∗∗ (p < 0.01), ∗∗∗ (p < 0.001).

47

Table 3. Conditional Logit by Income Group

Income Income Income<$50,000 ≥$50,000 & ≥$100,000

<$100,000

Price -0.561∗∗∗ -0.561∗∗∗ -0.459∗∗∗ -0.458∗∗∗ -0.363∗∗∗ -0.363∗∗∗(0.021) (0.020) (0.011) (0.011) (0.010) (0.010)

Elect. Cost -0.315∗ -0.327∗∗ -0.562∗∗∗ -0.563∗∗∗ -0.988∗∗∗ -0.989∗∗∗(0.126) (0.126) (0.170) (0.169) (0.251) (0.251)

Rebate 0.063∗∗ 0.073∗∗∗ 0.020 0.020 -0.025 -0.023(0.022) (0.021) (0.017) (0.016) (0.018) (0.018)

ENERGY STAR 0.072∗∗∗ -0.107 0.087∗∗∗ -0.273∗∗∗ 0.134∗∗∗ 0.015(0.018) (0.0607) (0.020) (0.051) (0.021) (0.058)

ENERGY STARX Educ College -0.080∗∗∗ -0.010 -0.004

(0.023) (0.017) (0.018)X Educ Graduate 0.151∗∗∗ 0.082∗∗ 0.028

(0.042) (0.026) (0.021)X Income +$16.5K 0.165∗∗∗ 0.086∗∗∗ 0.032

(0.030) (0.025) (0.020)X Income +$33K 0.184∗∗∗ 0.210∗∗∗ -0.007

(0.026) (0.022) (0.019)X Age > 30,≤ 45 0.096∗ 0.116∗∗ 0.093∗

(0.046) (0.039) (0.042)X Age > 45,≤ 55 0.012 0.039 0.051

(0.050) (0.037) (0.043)X Age > 55,≤ 70 0.146∗∗ 0.136∗∗∗ 0.058

(0.045) (0.038) (0.043)X Age > 70 0.050 0.136∗∗∗ -0.017

(0.045) (0.041) (0.047)X FamSize 2 0.183∗∗∗ 0.231∗∗∗ 0.130∗∗∗

(0.034) (0.026) (0.033)X FamSize 3 or 4 0.199∗∗∗ 0.247∗∗∗ 0.117∗∗∗

(0.034) (0.026) (0.031)X FamSize 5 or more 0.220∗∗∗ 0.200∗∗∗ 0.086∗∗

(0.038) (0.027) (0.033)X Political Democrat -0.157∗∗∗ -0.101∗∗∗ -0.059∗

(0.035) (0.024) (0.023)X Political Others -0.175∗∗∗ -0.119∗∗∗ -0.066∗∗

(0.031) (0.022) (0.023)Interpretation

Price Elasticity -7.29 -7.54 -5.97 -5.97 -4.72 -4.72

WTP ES τ/η ($) 12.7 varies with demo. 18.9 varies with demo. 36.9 varies with demo.

Prob. Take Rebate ψ/η 0.11 0.13 0.04 0.04 -0.07 -0.07

Implied Discount Rate 1.78 1.82 0.82 0.82 0.37 0.37Nb Obs. 49,279 76,115 76,115Nb Clusters 922 922 922Note: Standard errors clustered at the zipcode level in parentheses: ∗ (p < 0.05),∗∗ (p < 0.01), ∗∗∗ (p < 0.001).

48

Table 4. Conditional Logit by Income Group, contd.

Income Income Income<$50,000 ≥$50,000 & ≥$100,000

<$100,000

Price -0.561∗∗∗ -0.459∗∗∗ -0.363∗∗∗(0.020) (0.011) (0.010)

Elect. Cost 0.045 -0.452 0.857∗(0.297) (0.274) (0.351)

Rebate 0.073∗∗∗ 0.021 -0.022(0.021) (0.016) (0.018)

ENERGY STAR -0.090 -0.263∗∗∗ 0.050(0.060) (0.051) (0.058)

Interactions: Elect. Cost ENERGY Elect. Cost ENERGY Elect. Cost ENERGYSTAR STAR STAR

X Educ College -0.038 -0.081∗∗∗ -0.044 -0.011 -0.284∗∗∗ -0.009(0.093) (0.023) (0.079) (0.017) (0.086) (0.018)

X Educ Graduate 0.012 0.149∗∗∗ -0.377∗∗ 0.075∗∗ -0.454∗∗∗ 0.021(0.201) (0.043) (0.120) (0.027) (0.113) (0.021)

X Income +$16.5K 0.232 0.169∗∗∗ 0.458∗∗∗ 0.095∗∗∗ 0.001 0.032(0.121) (0.030) (0.114) (0.025) (0.099) (0.020)

X Income +$33K 0.404∗∗∗ 0.190∗∗∗ 0.460∗∗∗ 0.219∗∗∗ -0.214∗ -0.011(0.103) (0.026) (0.100) (0.022) (0.095) (0.019)

X Age > 30,≤ 45 0.036 0.098∗ -0.093 0.115∗∗ -0.312 0.084∗(0.176) (0.047) (0.153) (0.039) (0.195) (0.043)

X Age > 45,≤ 55 -0.867∗∗∗ -0.008 -0.912∗∗∗ 0.015 -1.078∗∗∗ 0.024(0.179) (0.050) (0.156) (0.037) (0.203) (0.044)

X Age > 55,≤ 70 -2.081∗∗∗ 0.110∗ -2.356∗∗∗ 0.088∗ -2.391∗∗∗ 0.014(0.175) (0.045) (0.164) (0.038) (0.203) (0.044)

X Age > 70 -3.919∗∗∗ 0.002 -3.813∗∗∗ 0.075 -3.781∗∗∗ -0.070(0.188) (0.045) (0.189) (0.042) (0.240) (0.047)

X FamSize 2 0.609∗∗∗ 0.192∗∗∗ 0.565∗∗∗ 0.242∗∗∗ -0.266 0.127∗∗∗(0.158) (0.034) (0.129) (0.027) (0.175) (0.033)

X FamSize 3 or 4 1.084∗∗∗ 0.213∗∗∗ 0.806∗∗∗ 0.257∗∗∗ -0.027 0.116∗∗∗(0.149) (0.034) (0.124) (0.026) (0.170) (0.030)

X FamSize 5 or more 1.389∗∗∗ 0.239∗∗∗ 1.098∗∗∗ 0.221∗∗∗ -0.016 0.085∗∗(0.169) (0.038) (0.139) (0.027) (0.179) (0.033)

X Political Democrat -0.145 -0.160∗∗∗ -0.023 -0.102∗∗∗ -0.195 -0.062∗∗(0.160) (0.035) (0.119) (0.024) (0.120) (0.023)

X Political Others -0.102 -0.178∗∗∗ 0.339∗∗ -0.113∗∗∗ 0.200 -0.063∗∗(0.148) (0.032) (0.114) (0.022) (0.123) (0.023)

Nb Obs. 49,279 76,115 76,115Nb Clusters 922 922 922Notes: Standard errors clustered at the store level. ∗ (p < 0.05), ∗∗ (p < 0.01),∗∗∗ (p < 0.001)

49

Table 5. Information Acquisition Model: 2008

Income Income Income<$50,000 ≥$50,000 & ≥$100,000

<$100,000

Behavioral Parameters Purchase Decision:Price (η) -0.543∗∗∗ (0.001) -0.448∗∗∗ (0.001) -0.357∗∗∗ (0.001)ENERGY STAR τ I -0.312∗∗∗ (0.065) 0.008 (0.020) -0.064∗ (0.030)ENERGY STAR τES 0.825∗∗∗ (0.123) 0.429∗∗∗ (0.044) 0.513∗∗∗ (0.054)Rebate (ψ) 0.282∗∗ (0.112) 0.132∗∗ (0.047) 0.054 (0.043)Elec. Costs (θ) -5.197∗∗∗ (0.407) -3.359∗∗∗ (0.147) -3.861∗∗∗ (0.224)Kh -11.116∗∗∗ (1.245) -8.124∗∗∗ (1.634) -6.041∗∗∗ (1.548)Km -2.368∗∗ (0.846) -3.928∗ (1.864) -1.336 (1.544)Educ College (βI) 0.155 (0.189) 0.273 (0.278) 0.288 (0.235)Educ Graduate (βI) 0.205 (0.426) 1.941∗∗∗ (0.469) 0.578 (0.262)FamSize (βI) -0.325∗∗∗ (0.060) -0.477∗∗∗ (0.077) -0.180∗∗ (0.056)Age (βI) 0.185∗∗∗ (0.018) 0.207∗∗∗ (0.020) 0.166∗∗∗ (0.015)Political Democrat (βI) -0.560∗ (0.332) -0.958∗∗ (0.409) -0.785∗∗ (0.333)Political Others (βI) -0.490 (0.328) -1.992∗∗∗ (0.420) -1.057∗∗ (0.329)Educ College (βES) -0.052 (0.268) 0.695∗ (0.359) 0.301 (0.256)Educ Graduate (βES) 1.439∗∗∗ (0.369) 2.477∗∗∗ (0.502 ) 0.405 (0.291)FamSize (βES) 0.081 (0.064) -0.020 (0.088) -0.166∗∗ (0.061)Age (βES) 0.042∗∗∗ (0.009) 0.061∗∗∗ (0.015) 0.056∗∗∗ (0.012)Political Democrat (βES) -1.310∗∗∗ (0.332) -1.368∗∗ (0.457) -1.235∗∗∗ (0.333)Political Others (βES) -1.581∗∗∗ (0.333) -2.041∗∗∗ (0.449) -1.313∗∗∗ (0.318)mean-ElecCost -2.663∗∗ (0.969) 3.209∗∗ (1.417) -1.294 (1.066)var-ElecCost -3.426 (4.983) -30.029∗∗∗ (8.911) 3.275 (6.569)Nb Models (γI) 0.013 (0.008) 0.004 (0.005) 0.005 (0.003)Variance Price (γI) 0.034 (0.061) -0.185 (0.135 ) -0.040 (0.131)Variance FE (γI) -0.197 (0.125) 0.156 (0.384) -0.603 (0.463)Proportion-Estar 2.667∗∗∗ (0.805) 6.817∗∗∗ (1.350) 2.677∗∗∗ (0.702)Nb Models (γES) -0.090∗∗∗ (0.017) -0.048∗∗∗ (0.009) -0.024∗∗∗ (0.005)Variance Price (γES) 0.067 (0.066) 0.245 (0.161) 0.127 (0.126)Variance FE (γES) -0.089 (0.135) -0.293 (0.454) -0.255 (0.432)

InterpretationOwn-Price Elasticity -7.1 -5.8 -4.6Implicit Discount Rate 0.08 0.11 0.06WTP ES Label, e = I ($) -57.4 1.9 -18.0WTP ES Label, e = ES ($) 151.8 95.8 143.8Prob. Taking Rebate 0.52 0.29 0.15Q(e = I) 0.19 0.34 0.30Q(e = ES) 0.16 0.24 0.37Q(e = U) 0.65 0.41 0.33

Nb Obs. 49,279 76,115 76,115Notes: In the estimation, estimated product fixed effects enter the individual choice probabilities,and are treated as data. Standard errors are obtained using the Murphy and Topel’s approach(1985). ∗ (p < 0.05), ∗∗ (p < 0.01), ∗∗∗ (p < 0.001)

50

Table 6. The Opportunity Cost of Imperfect Energy Information

Perfectly Informed Perfectly Informed Perfectly Informedvs vs vs

Information Acquisition Uninformed ENERGY STAR

Consumer Surplus with Label Effect ($)Income <$50,000 25 33 18≥$50,000 & <$100,000 14 21 19≥$100,000 20 32 26All Income 19 28 21

Consumer Surplus without Label Effect ($)Income <$50,000 36 40 56≥$50,000 & <$100,000 15 20 28≥$100,000 26 32 46All Income 24 29 42

kWh/year purchased-18 -26 -24

Externality Costs ($)-14 -21 -19

Producer Surplus ($)-21 -22 -47

Welfare ($)with label 12 27 -7without label 17 28 14

Notes: In each column, a choice model is compared to a model where all consumers are perfectly informed. The firstcolumn compares the outcomes obtained with the information acquisition model. The second column compares achoice model where all consumers are uninformed. The third column compares a choice model where all consumersrely on ENERGY STAR. The table shows that the consumers surplus decreases in all cases when consumers are notperfectly informed, but profits increase. The externality costs increase under imperfect information.

51

Online Appendix

Appendix A. Comparative Statics Results

Information Acquisition Costs. Naturally, we should expect that consumers will collect and

process more energy information the lower the costs to do so. In particular, a consumer should

always choose to be fully informed if there are no extra costs. The present model is consistent with

this intuition, whether ENERGY STAR information is available or not.

Proposition 1. (i) Suppose that ENERGY STAR information is not available. If K(e =

U) = K(e = I), it is optimal for the consumer to select e = h.

(ii) Suppose that ENERGY STAR information is available. If K(e = U) = K(e = ES) = K(e =

I), it is optimal for the consumer to select e = I. Moreover, if K(e = U) = K(e = ES),

e = ES is strictly better than e = U for the consumer.

Proof. This is true by the fact that the expectation of the maximum of random variables is always

greater that the maximum of their expectations. In particular, consider some set of random vari-

ables {Y1, Y2, ..., Yk}. The distribution of max1≤j≤k{Yj} (first order) stochastically dominates the

distribution of Yl for any l ∈ {1, ..., k}. This implies that E[max1≤j≤k Yj ] ≥ E[Yl] for l = 1, ..., k,

and thus,

(14) E[ max1≤j≤k

Yj ] ≥ max1≤j≤k

E[Yj ].

To show (i), it suffices to show that

Eε,C

[maxj{Uij(δj , ηPj , Cj , εij)}

]≥ Eε

[maxj{EC [Uij(δj , ηPj , Cj , εij)]}

],

which implies that (i) holds for K = 0. I show a stronger inequality; in particular, that for any εij ,

EC

[maxj{Uij(δj , ηPj , Cj , εij)}

]≥[maxj{EC [Uij(δj , ηPj , Cj , εij)]}

].

This follows from (14) if I set

Yj ≡ Uij(δj , ηPj , Cj , εij).

52

This concludes the proof for (i).

The proof for (ii) is similar.

Crowding-Out Effect. A simple, but important implication to the above result is that if the costs

of processing and collecting ENERGY STAR information are lower than the costs of searching for

energy costs, some consumers may prefer to select the maximum level of effort than to not collect

information at all, but could prefer a medium level of effort than a maximum one. Formally,

Corollary 1. If K(e = ES) < K(e = I), then for some consumers

V(e = U) < V(e = I) < V(e = ES)

Proof. The proof follows directly from Proposition 1.

This formally shows that the ENERGY STAR certification induces some consumers to be less

informed and crowds out efforts to fully account for energy costs.

Uncertainty about Energy Costs. The present model stipulates that uncertainty in beliefs is

the main driver that induces consumers to search for energy information. Therefore, the model

should predict that the larger the uncertainty in beliefs, the more likely consumers are to search.

I now show this result. I focus on the case that consumers’ beliefs become more uncertain, but

remain unbiased. This scenario is notably consistent with Allcott (2011)’s findings that shows that

consumers’ beliefs about future gas prices are on average unbiased, but largely uncertain.

Consider the following definition. Beliefs about X represented by a distribution F are more

uncertain that beliefs F if F second order stochastically dominates F :

(15)∫ b

xF (x)dx ≥

∫ b

xF (x)dx

for all b.

53

Proposition 2. If Xj ∼ F ,∀j and Xj ∼ F,∀j, with∫ bx F (x)dx ≥

∫ bx F (x)dx for all b, then

Eε,X ,S

[maxj{Uij(δj , ηPj , Xj(S), d, εij)}|I(e)

]≥ Eε,X ,S

[maxj{Uij(δj , ηPj , Xj(S), d, εij)}|I(e)

]

Proof. First note that by the definition of second order stochastically dominance, if Xj second-order

stochastically dominates Xj , and if Xj and Xj have the same mean, then E[h(Xj)] ≥ E[h(Xj)] for

all concave function h. Given that the maximum is a concave function, I then have E[h(Xj , Y )] ≥

E[h(Xj , Y )] for any variable Y .

Proposition 2 simply says that the larger the variance in energy costs, the higher is the value

of information. This also implies that ENERGY STAR will lead to more sub-optimal choices, in

expectation, in choice sets where products are largely disperse in the energy efficiency characteristics

space.

54

Appendix B. Data Cleaning and Manipulation

Creating a Random Sample. To perform the estimation of the demand model, a random sub-

sample of the transactions is used. The sub-sample is constructed as follow.

First, the sub-sample is drawn from the set of transactions that fit the following criteria (the

restricted sample):

• transactions made by consumers that are homeowners;

• transactions made by consumers living in single family housing units; and

• transactions made by consumers that made no more that one refrigerator purchase in any

given year.

Second, the following stratified sampling method is used to create the sub-sample. For a given

targeted sample size, I sample transactions for three different income groups:

• households with income of less than ≥$50,000;

• households with income between $50,000 and $100,000;

• and households with income of more than $100,000.

Average Electricity Prices. The use of average electricity prices is partly motivated by recent

empirical evidence (Borenstein (2010), Ito (2012)) that suggests that electricity consumers may in

fact respond to variation in average prices, more than marginal prices. In the present case, the use

of average electricity prices is also dictated by the fact that household’s location is not perfectly

known. Therefore, it is impossible to match households with their exact electricity tariff and infer

marginal price.

Average electricity prices at the county level are computed as follow. Using form EIA-861 of

the Energy Information Administration, I compute the average residential electric price for each

electric utility operating in the US for the years 2008. I then match electric utility territories with

each of the county where I sampled at least one store. For counties with only one electric utility, I

55

use the average electricity price for this particular utility. For counties with several electric utilities,

I take the arithmetic mean of each utility average price to construct the county level price.

Appendix C. Alternative Estimators

The Average Consumer. Adding an outside option to the conditional logit, I obtain a linear

expression for the market shares in region r at time t (Berry 1994):

(16) ln(qjrt) = τDjt − ηPjrt + ψRrtXDjt − θCjr + γj + αr + αt + ζjrt

where qjrt is the quantity of refrigerator model j sold during week t in store r, and ζjrt is a market-

specific unobservable. Not all refrigerator models sell every week in every store, the dependent

variable thus takes the value zero for a large number of observations. Equation 16 is then estimated

with a negative binomial model. The sample used for the estimation consists of all transactions

observed during the period 2007-2009. Stores with a low number of sales were excluded, which

left 545 stores for the estimation. Price time series for each refrigerator model are week and store

specific. The choice set are store and month specific, and constructed using the same methodology

than for the conditional logit model.

In the first specification, product, week, and store fixed effects are included, and the electricity

prices averaged at the county level are used. The estimate of the price coefficient corresponds to an

own-price elasticity of -2.89, which is about half the elasticity obtained with the conditional logit

model. The present price elasticity has a different interpretation here given that an outside option

has been added, and can be interpreted as being long-run. The size of the label effect is 0.041, and

corresponds to a WTP for the ENERGY STAR label of 18$. This replicates closely the estimate

obtained with the condition logit (19$). The effect of rebate is negative and not significant. The

estimate of the coefficient on electricity costs is negative and significant, and implies a discount

rate of 41%, which is lower than for the conditional logit model (62%).

The second specification replace the week fixed effects, with brand-week fixed effects. Like in

the conditional logit model, it has few effects on the price coefficient, suggesting that marketing

56

efforts are not an important source of bias. The coefficient for the ENERGY STAR label is larger

under this specification.

The third specification uses electricity prices averaged at the state level. This impacts the

coefficient on electricity costs. The coefficient still negative and significant, but implies a discount

rate of 29%. Similar than in the condition logit model, measurement error in how electricity prices

are measured has an important effect.

The fourth specification includes store-year fixed effects. This provides a better control for any

region specific unobservables. Doing so has few impact on the coefficients.

Interaction with Demographics. Table 8 presents the results from a conditional logit model

where the coefficient on prices is interacted with demographic information, in addition of the

coefficients on electricity costs, and ENERGY STAR dummy. The results complement Table 4,

and show that the main patterns hold.

57

Table 7. Negative Binomial Model

(I) (II) (III) (IV)

Price (η) -0.222∗∗∗ -0.219∗∗∗ -.219∗∗∗ -0.219∗∗∗(0.006) (0.006) (.006) (0.006)

ENERGY STAR (τ) 0.041∗∗∗ 0.076∗∗∗ .075∗∗∗ 0.077∗∗∗(0.009) (0.010) (.010) (0.010)

Rebate (ψ) -0.011 -0.012 -.010 -0.013(0.013) (.013) (.013) (0.014)

Elect. Cost (θ) -.532∗∗∗ -0.533∗∗∗ -.740∗∗∗ -0.543∗∗∗(0.115) (.115) (0.088) (0.120)

Product FE Yes Yes Yes Yes

ZipCode FE Yes Yes Yes Yes

Week FE Yes No No No

BrandXWeek FE No Yes Yes Yes

Avg. Elec. Price County County State County

Observations 9.50e+06 9.50e+06 9.50e+06 9.50e+06Nb Clusters 545 545 545 545

Interpretation: TO UPDATEPrice Elasticity -2.89 -2.85 -2.85 -2.85WTP ES τ/η 18.0 34.7 34.2 35.2Prob. Take Rebate ψ/η -0.05 -0.05 -0.05 -0.06Implied Discount Rebate 0.417 0.410 0.293 0.402Note: † (p¡0.10) ∗ (p < 0.05), ∗∗ (p < 0.01), ∗∗∗ (p < 0.001)

58

Table 8. Conditional Logit: Interaction with Demographics

Income Income Income<$50,000 ≥$50,000 & ≥$100,000

<$100,000

Price -0.586∗∗∗ -0.496∗∗∗ -0.404∗∗∗(0.022) (0.013) (0.012)

Elect. Cost 0.445 0.222 1.643∗∗∗(0.320) (0.287) (0.361)

Rebate 0.073∗∗∗ 0.020 -0.022(0.021) (0.016) (0.018)

ENERGY STAR 0.0179 -0.051 0.313∗∗∗(0.074) (0.063) (0.069)

Interactions: Elect. ENERGY Price Elect. ENERGY Price Elect. ENERGY PriceCost STAR Cost STAR Cost STAR

X Educ College 0.011 -0.066∗ -0.003 -0.076 -0.020 0.001 -0.473∗∗∗ -0.065∗∗ 0.008∗∗∗(0.101) (0.027) (0.003) (0.086) (0.021) (0.002) (0.092) (0.022) (0.002)

X Educ Grad -0.445 0.016 0.024∗∗∗ -0.733∗∗∗ -0.030 0.017∗∗∗ -0.737∗∗∗ -0.060∗ 0.012∗∗∗(0.230) (0.051) (0.005) (0.129) (0.031) (0.003) (0.121) (0.025) (0.002)

X Inc +$16.5K -0.030 0.088∗ 0.015∗∗∗ 0.255∗ 0.037 0.010∗∗∗ -0.308∗∗ -0.059∗ 0.013∗∗∗(0.138) (0.036) (0.004) (0.124) (0.030) (0.003) (0.106) (0.024) (0.002)

X Inc +$33K 0.215 0.132∗∗∗ 0.011∗∗ -0.072 0.055∗ 0.027∗∗∗ -0.759∗∗∗ -0.167∗∗∗ 0.023∗∗∗(0.116) (0.030) (0.003) (0.111) (0.026) (0.002) (0.107) (0.021) (0.002)

X Age >30,≤45 -0.350 -0.034 0.024∗∗∗ -0.444∗∗ -0.010 0.020∗∗∗ -0.826∗∗∗ -0.089 0.025∗∗∗(0.197) (0.056) (0.006) (0.161) (0.044) (0.004) (0.201) (0.051) (0.004)

X Age >45,≤55 -1.114∗∗∗ -0.088 0.014∗ -1.053∗∗∗ -0.034 0.008∗ -1.366∗∗∗ -0.077 0.015∗∗∗(0.198) (0.058) (0.006) (0.164) (0.043) (0.004) (0.208) (0.052) (0.004)

X Age >55,≤70 -2.304∗∗∗ 0.042 0.013∗ -2.592∗∗∗ 0.0010 0.013∗∗∗ -2.520∗∗∗ -0.043 0.008∗(0.194) (0.053) (0.005) (0.174) (0.044) (0.004) (0.208) (0.053) (0.004)

X Age >70 -3.392∗∗∗ 0.137∗ -0.026∗∗∗ -3.425∗∗∗ 0.164∗∗∗ -0.015∗∗∗ -3.127∗∗∗ 0.072 -0.021∗∗∗(0.206) (0.054) (0.006) (0.196) (0.047) (0.004) (0.240) (0.057) (0.005)

X FamSize 2 0.342 0.109∗∗ 0.015∗∗ 0.331∗ 0.168∗∗∗ 0.012∗∗∗ -0.588∗∗ 0.025 0.015∗∗∗(0.177) (0.042) (0.005) (0.139) (0.032) (0.003) (0.179) (0.039) (0.003)

X FamSize 3-4 0.628∗∗∗ 0.074 0.026∗∗∗ 0.436∗∗ 0.140∗∗∗ 0.019∗∗∗ -0.480∗∗ -0.021 0.020∗∗∗(0.168) (0.041) (0.004) (0.136) (0.032) (0.003) (0.175) (0.037) (0.003)

X FamSize 5 + 0.817∗∗∗ 0.060 0.033∗∗∗ 0.750∗∗∗ 0.112∗∗∗ 0.018∗∗∗ -0.364∗ -0.021 0.016∗∗∗(0.183) (0.046) (0.005) (0.152) (0.033) (0.003) (0.183) (0.038) (0.003)

X Pol Dem. 0.080 -0.098∗ -0.011∗ 0.224 -0.030 -0.012∗∗∗ 0.047 0.004 -0.009∗∗∗(0.187) (0.042) (0.005) (0.130) (0.028) (0.002) (0.129) (0.026) (0.002)

X Pol Others 0.150 -0.106∗∗ -0.013∗∗ 0.558∗∗∗ -0.049 -0.011∗∗∗ 0.413∗∗ -0.003 -0.009∗∗∗(0.174) (0.039) (0.004) (0.123) (0.027) (0.002) (0.132) (0.025) (0.002)

Notes: Standard errors clustered at the store level. ∗ (p < 0.05), ∗∗ (p < 0.01),∗∗∗ (p < 0.001)

59

Appendix D. FKRB’s Estimator

The FKRB’s estimator models the CDF of the random parameters as a mixture of point masses.

Alternatively, FKRB propose to estimate a smooth density by modeling the distribution of the

parameters as a mixture of normal densities. This requires to specify M normal basis functions

with a predetermined mean and variance, and use simulation to construct the estimator.

In particular, the mth basis function is defined as the product of the marginals of the K random

parameters:

(17) N(β|µm, σm) =K∏k=1

N(β|µmk , σmk )

The simulated choice probability is thus given by:

(18) Qijrt ≈M∑m

αm

(1S

S∑s=1

Pmijrt|βm,s

)

where βm,s is the sth draw from the rth normal basis. The estimate of α is thus given by:

(19) α = argminα1NJ

I∑i=1

J∑j=1

(yijrt −

M∑m

αm

(1S

S∑s=1

Pmijrt|βm,s

))

s.t.M∑m

αm = 1, αm ≥ 0

For the present application, η, ψ and γj are taken as data, and they are fixed at the MLE

estimates (Model 1, Table 2). The joint density of the parameters θ and τ is a mixture of M = 108

normal basis functions. To construct the basis functions, 9 marginals for the parameter θ, and 11

marginals for the parameter τ are used. The means of the marginals are defined relative to the MLE

estimates of the coefficients τ and θ. In particular, the means of the marginals for the parameter θ

is the vector: [2.25θ, 2θ, 1.75θ, 1.25θ, θ, 0.75θ, 0.5θ, 0.25θ, 0.05θ]. The means of the marginals for the

parameter τ is the vector: [8τ , 6τ , 4τ , 2.5τ , 1.75τ , 1.25τ , τ , 0.75τ , 0.5τ , 0.25τ , 0]. For each marginal,

60

the standard deviation is set to a small value corresponding to 0.1% of the mean of the marginal.

The estimation is carried with the matlab package lsqlin.

61

Appendix E. Policy Analysis

E.1. Emission Factors

Table 9. Emission Factors and Externality Costs

Non-baseload Output Emission Rates (U.S. Average)Pollutant Estimate SourceCO2 1,583 lb/MWhCH4a 35.8 lb/GWhN2Oa 19.9 lb/GWh USEPA, eGRID2007SO2 6.13 lb/MWhNOx 2.21 lb/MWh

Damage Cost (2008 $)Pollutant Low Estimate High Estimate SourceCO2 $21.8/t $67.1/t Greenstone, Kopits and Wolverton (2011)SO2 $2,060/t $6,700/t low: Muller and Mendelsohn (2012), high: USEPAb

NOx $380/t $4,591/t low: Muller and Mendelsohn (2012), high: DOEc

Notes: (a) Externality costs associated to CH4 and N2O are assumed to be the same than for CO2. CH4 andN2O are converted in CO2 equivalent using estimates of global warming potential (GWP). The GWP used for CH4is 25, and the GWP used for N2O is 298. Source: IPCC Fourth Assessment Report: Climate Change 2007. (b)Estimate used in the illustrative analysis of the 2012 regulatory impact analysis for the proposed standards for electricutility generating units. (c) Higher value of the estimate used in the Federal Rule for new minimum energy-efficiencystandards for refrigerators (1904-AB79).

62

E.2. Sensitivity Tests

Table 10. The Opportunity Cost of Imperfect Energy Information: LowElectricity Costs, No Rebate

Perfectly Informed Perfectly Informed Perfectly Informedvs vs vs

Information Acquisition Uninformed ENERGY STAR

Consumer Surplus with Label Effect ($)Income <$50,000 13 18 5≥$50,000 & <$100,000 8 12 12≥$100,000 11 18 13All Income 10 16 11

Consumer Surplus without Label Effect ($)Income <$50,000 20 24 32≥$50,000 & <$100,000 9 12 20≥$100,000 16 18 30All Income 14 17 27

kWh/year purchased-14 -21 -19

Externality Costs ($)-11 -17 -15

Producer Surplus ($)-21 -22 -47

Welfare ($)with label 4 15 -15without label 8 16 1

Notes: In each column, a choice model is compared to a model where all consumers are perfectly informed.The first column compares the outcomes obtained with the information acquisition model. The secondcolumn compares a choice model where all consumers are uninformed. The third column compares achoice model where all consumers rely on ENERGY STAR. The table shows that the consumers surplusdecreases in all cases when consumers are not perfectly informed, but profits increase. The externalitycosts increase under imperfect information.

63

Table 11. The Opportunity Cost of Imperfect Energy Information: HighElectricity Costs, $50 Rebate

Perfectly Informed Perfectly Informed Perfectly Informedvs vs vs

Information Acquisition Uninformed ENERGY STAR

Consumer Surplus with Label Effect ($)Income <$50,000 52 63 41≥$50,000 & <$100,000 31 44 39≥$100,000 41 63 52All Income 40 56 45

Consumer Surplus without Label Effect ($)Income <$50,000 60 67 72≥$50,000 & <$100,000 31 42 49≥$100,000 47 61 72All Income 44 55 63

kWh/year purchased-27 -37 -34

Externality Costs ($)-11 -17 -15

Producer Surplus ($)-30 -30 -59

Welfare ($)with label 21 42 0without label 25 42 19

Notes: In each column, a choice model is compared to a model where all consumers are perfectly informed.The first column compares the outcomes obtained with the information acquisition model. The secondcolumn compares a choice model where all consumers are uninformed. The third column compares achoice model where all consumers rely on ENERGY STAR. The table shows that the consumers surplusdecreases in all cases when consumers are not perfectly informed, but profits increase. The externalitycosts increase under imperfect information.

64

Table 12. The Opportunity Cost of Imperfect Energy Information: NoBunching at Previous Certification Requirement

Perfectly Informed Perfectly Informed Perfectly Informedvs vs vs

Information Acquisition Uninformed ENERGY STAR

Consumer Surplus with Label Effect ($)Income <$50,000 9 24 -20≥$50,000 & <$100,000 6 15 3≥$100,000 8 24 4All Income 8 21 -2

Consumer Surplus without Label Effect ($)Income <$50,000 21 28 17≥$50,000 & <$100,000 8 11 11≥$100,000 14 19 20All Income 14 18 16

kWh/year purchased-15 -24 -15

Externality Costs ($)-12 -19 -12

Producer Surplus($)-20 -11 -41

Welfare ($)with label -1 29 -32without label 5 26 -14

Notes: In each column, a choice model is compared to a model where all consumers are perfectly informed.The first column compares the outcomes obtained with the information acquisition model. The secondcolumn compares a choice model where all consumers are uninformed. The third column compares achoice model where all consumers rely on ENERGY STAR. The table shows that the consumers surplusdecreases in all cases when consumers are not perfectly informed, but profits increase. The externalitycosts increase under imperfect information.